A UAV Small Object Detection Method Based on Multi-Branch Detail Enhancement and Cross-Scale Attention Fusion

Abstract

In unmanned aerial vehicle (UAV)-based small object detection, enhancing object perception under complex backgrounds remains a critical challenge for current detection models. Owing to the small scale, low pixel occupancy, and cluttered backgrounds of UAV small objects, their discriminative features are prone to attenuation in deep networks, which limits multi-scale feature fusion and consequently increases detection difficulty. To resolve these issues, this paper presents a Detail-Aware Multi-scale Fusion Network (DMF-Net) for UAV small object detection, consisting of two core modules: the Multi-Branch Detail-Enhanced Module (MBDE) and the Dual-Attention Fusion (DAF) Module. First, during feature extraction, a multi-branch detail enhancement module with serial convolutions in each branch and residual connections is introduced to strengthen high-frequency details and local textures while preserving semantic consistency. Second, at the feature fusion stage, a dual-attention-based feature fusion module with a symmetric interaction structure is designed to dynamically evaluate the significance of features at different scales via adaptive attention mechanisms, enabling symmetric cross-scale interaction and fine-grained feature complementarity. The experimental results obtained on the challenging VisDrone and TinyPerson datasets confirm that DMF-Net outperforms existing state-of-the-art detection methods in terms of the accuracy of small object detection, while maintaining high inference efficiency. On the TinyPerson dataset, DMF-Net improves AP by 1.4%, AP_50 by 4.1%, and AP_75 by 0.6% compared with YOLO11n, while maintaining 97.2 FPS. Furthermore, it shows promising performance in complex backgrounds and densely populated scenarios.

1. Introduction

Owing to the swift technological progress in unmanned aerial vehicle (UAV), object detection based on UAV platforms has demonstrated substantial application value in military surveillance, public security, film production, and logistics delivery. Compared to general object detection, objects in UAV imagery typically exhibit characteristics such as small scale, low pixel occupancy, and complex backgrounds, making small objects more susceptible to interference from environmental textures and noise during imaging, thereby posing greater challenges for detection. Consequently, the urgent need to tackle the enhancement of feature extraction and fusion for small objects under complex backdrops in UAV object detection has become a top priority.

In recent years, a growing body of study has focused on detecting small objects using UAV, yielding notable advancements in this domain. It is evident that, in accordance with divergent detection paradigms, extant methodologies may be categorised into two-stage and one-stage approaches. Two-stage methods, such as Faster R-CNN [1] and Cascade R-CNN [2], achieve high detection accuracy through a two-step process of region proposal generation followed by classification and regression. Nevertheless, when it comes to detecting small objects in complicated scenes, these approaches often have limited inference speeds and lack the necessary discriminative features for small objects. In contrast, one-stage methods, including the YOLO [3,4,5,6] series and SSD [7], perform direct object prediction on multi-scale feature maps, enabling efficient end-to-end inference. Among them, the YOLO series progressively enhances feature representation by employing serial linear convolutions, residual connections, and multi-scale feature fusion, thereby improving small-object detection performance while preserving fast inference speed. Nevertheless, constrained by the local nature of convolutional operations, CNNs remain limited in capturing global contextual relationships in complex scenes. With the introduction of Transformer architectures, DETR [8] performs object localization via end-to-end sequence modeling and a query-based mechanism [9], eliminating the need for predefined anchors and complex post-processing, and achieving notable improvements in detection accuracy. Despite these advantages, DETR exhibits slow convergence in small-object detection and struggles to extract detailed local features. To overcome these issues, RT-DETR [10] and D-Fine [11] accelerate convergence and improve small object detection performance by introducing efficient query mechanisms and feature enhancement strategies. Furthermore, DEIM [12] improves training efficiency and detection performance by refining the Hungarian matching strategy and sample assignment mechanism. In addition, Mamba [13], based on state-space models, enhances contextual modeling capability, providing a new approach to detecting small objects in complex scenes.

Although progress has been made in designing the detection framework and improving overall performance using existing methods, there are still multiple challenges in UAV-based small object detection [14]. First, current feature extraction processes largely rely on single propagation paths or serial linear convolutional structures. While such designs facilitate the progressive accumulation of high-level semantic information, they are insufficient in modeling high-frequency details and local textures that are critical for discriminating small objects, causing important cues to be gradually attenuated in deeply stacked convolutions [15]. Second, mainstream multi-scale feature fusion strategies typically adopt simple aggregation or fixed-weight mechanisms, lacking adaptive weighting of features across different scales [16]. This limitation hinders flexible cross-scale interaction in response to variations in object scale and contextual information, thereby restricting the effective utilization of fine-grained feature complementarity. Moreover, under UAV viewpoints with complex backgrounds or densely distributed targets, existing fusion approaches remain inadequate in jointly modeling local discriminative features and global semantic information. Overall, relying solely on conventional serial convolutional stacking and static fusion strategies is no longer sufficient to continuously improve UAV small object detection performance. Therefore, more targeted improvements are required in both feature extraction architecture design and cross-scale attention-based fusion mechanisms.

To solve the above issues, the present paper puts forward a proposal for a Detail-Aware Multi-scale Fusion Network (DMF-Net) from two aspects: feature extraction architecture and cross-scale attention-based feature fusion. First, at the feature extraction stage, a convolutional structure named the Multi-Branch Detail-Enhanced Module (MBDE) is introduced to enhance local perception. By adopting a multi-branch design, MBDE strengthens the extraction of high-frequency details and local texture information, while residual connections are incorporated to preserve the original semantic information. This design mitigates the gradual attenuation of discriminative cues for small objects during deep convolution and enhances the extraction of local features critical for their detection. Second, a multi-scale attention fusion module termed the Dual-Attention Fusion Module (DAF) is designed at the multi-scale feature fusion stage. By using attention weights to reflect the features’ significance at different scales, the proposed module adaptively adjusts the intensity of cross-scale information interaction, thereby enabling more effective fine-grained feature complementarity. These improvements enable the proposed method to extract local details and integrate global semantic information more effectively under complex backgrounds and in densely populated target scenarios. This provides more robust feature representations for small object detection by UAVs.

As illustrated in Figure 1, the image being inputted exhibits small-scale objects, dense arrangements, and complex backgrounds. In Figure 1a, the YOLO11n backbone can extract some semantic information, but its feature responses are weak and scattered across the image. During the subsequent feature fusion process in the neck, these limited discriminative cues are slightly enhanced, leading to more localized feature responses; however, the final detection results remain unsatisfactory. In contrast, as illustrated in Figure 1b, DMF-Net exhibits more concentrated and continuous feature responses for small objects across the backbone, neck, and head stages, which is particularly evident in the red-circled regions of the detection results. By strengthening fine-grained feature representations and improving the effective propagation of multi-scale information, the proposed model better preserves key semantic cues of small objects under complex backgrounds. This enables more reliable localization and detection of small objects, effectively alleviating the limitations of existing YOLO-series methods in aerial small object detection scenarios. The main technical contributions of this paper are outlined below:

1. In order to tackle the challenges of small object scales, attenuated discriminative features, and insufficient multi-scale feature fusion in UAV scenarios, this paper proposes DMF-Net, which is jointly composed of the MBDE and DAF modules. Experimental results show that DMF-Net achieves consistent and significant performance improvements across multiple UAV small object datasets. While maintaining high inference efficiency, it effectively enhances small object detection accuracy, showing strong robustness and generalization capability.

2. An enhanced feature extraction structure for UAV small object detection, termed MBDE, is designed. To address the gradual attenuation of local discriminative information caused by deep serial convolutions in small object detection, multi-branch convolution structures (MSDC-Conv) are introduced progressively to capture multi-scale local details. Residual connections are incorporated to preserve the origin of semantic information, thus enhancing the extraction of high-frequency details and local texture features at different depths while maintaining semantic consistency. The feature representation of small objects is improved by this design.

3. A cross-scale attention-based feature fusion mechanism, termed DAF, is designed. To address the limitations of existing multi-scale fusion methods, which typically rely on linear addition or fixed weighting and struggle to fully leverage information across scales, an attention-based adaptive fusion strategy is introduced. This strategy dynamically adjusts the intensity of cross-scale information interaction according to the relative significance of features at different scales, enabling more effective fine-grained feature complementarity and boosting the network’s perceptual ability with regard to small objects.

The following sections are to be found in this paper: Section 2 offers an overview of prior studies on small object detection, multi-branch convolution strategies, and attention-based cross-layer feature enhancement techniques. Section 3 provides a comprehensive account of the proposed method, encompassing the overarching framework and the design of its key modules. Section 4 describes the experimental setup and provides a systematic analysis and discussion of the results. Finally, Section 5 summarizes the paper and explores potential avenues for future research.

2. Related Work

2.1. UAV Small Object Detection

With the rapid development of deep learning, small object detection has become a prominent research topic in computer vision. Earlier methods primarily employed two-stage algorithms, such as the R-CNN [17] series, which separate the detection process into region proposal generation and classification. While R-CNN achieves precise localization and recognition, its major drawbacks include high computational cost and long training time. Subsequent improvements have significantly enhanced the efficiency and performance of two-stage detectors, but they often increase computational complexity and do not fully resolve these limitations. To improve real-time performance, one-stage algorithms were developed. One-stage methods eliminate the need for region proposals by directly dividing the image into fixed regions and predicting object classes and locations in an end-to-end manner. This design enables faster computation and higher detection efficiency, making them suitable for real-time applications. Representative examples include YOLO [3,4,5,6] and SSD [7]. Although CNN-based methods capture substantial local detail, the limitations of convolutional kernel size and model structure restrict their ability to capture long-range dependencies and global contextual information. The Transformer architecture, presented by Vaswani et al. [18] in 2017, effectively addresses the challenge of modeling global information and has demonstrated strong potential across various vision tasks. Zhu et al. [19] addressed the slow convergence and limited local feature capture of the original DETR with Deformable DETR, which incorporates a deformable attention mechanism to focus on critical details more effectively. This design accelerates model convergence and improves detection performance for small objects.

The rapid advancement of general object detection has also driven continuous progress in UAV small object detection [20]. However, the imbalance between feature semantics and resolution in multi-scale detection remains a challenge. To improve small object detection accuracy, Lin et al. [21] presented the Feature Pyramid Network (FPN) with lateral connections in 2017. Subsequent studies have leveraged the FPN framework to make various contributions to UAV small object detection. For example, Zhao et al. [22] designed a UAV small object detection method based on YOLOv7, which optimized feature fusion structures and introduced attention mechanisms to address insufficient small object features and large multi-scale discrepancies in UAV scenes, thereby enhancing feature extraction performance and generalization capability. Li et al. [23] proposed an improved YOLOv7-UAV algorithm to tackle insufficient feature extraction and limited multi-scale adaptation in UAV images. By optimizing the prediction head to better capture small object features and reduce redundant computations, as well as improving multi-scale fusion in the feature pyramid, the method demonstrated notable enhancements in terms of both the precision of detection and the efficiency of inference. Xi et al. [24] introduced FiFoNet, a fine-grained object focusing network, which incorporates a fine-grained object focusing module, optimized multi-scale feature fusion strategies, and enhanced training methods. This approach effectively addresses large target scale variations, strong background interference, and insufficient fine-grained features in UAV imagery, achieving simultaneous improvements in detection accuracy, robustness, and real-time performance. Zhang et al. [25] proposed an attention-based deep neural network to overcome the limited ability of traditional sequential modeling to capture long-range dependencies. By introducing multi-head self-attention, the method enables parallel computation of global dependencies, significantly improving the model’s perception of contextual information. Wang et al. [26] developed a multi-modal fusion deep learning framework that incorporates cross-modal attention and adaptive feature alignment to bridge the semantic gap in heterogeneous data fusion, effectively mitigating limitations such as inconsistent feature distributions and weakened semantic associations in conventional methods.

Despite the success of deep learning algorithms in detecting small objects in UAVs, more precision is needed.

2.2. Multi-Branch Convolutions

Due to the inherent difficulties of small object detection—including sparse pixel coverage, limited content to capture, susceptibility to noise, and detail loss caused by downsampling—single-path convolutional structures are often insufficient for extracting meaningful small object features. Recent studies have shown that, compared to single-channel convolutions, parallel multi-channel convolution kernels can significantly improve feature extraction for small objects. Parallel heterogeneous convolutions offer advantages such as enlarged receptive fields, enhanced localization, improved detail capture, and increased robustness of feature representations.

Zhang et al. [27] introduced FFCA-YOLO, a detection framework enhanced with a dedicated Feature Enhancement Module (FEM). This module employs a parallel structure with multiple convolutional operations employing varied dilation factors, effectively expanding the receptive field of the convolutional layer without increasing the parameter count or reducing the feature map resolution. As a result, it can more fully capture multi-scale spatial information encompassing both fine-grained spatial features and high-level semantic context. Through efficiently integrating contextual features across scales, FEM improves the model’s detection capability and robustness for small-scale targets. Yang et al. [28] developed a multi-modal fusion deep learning framework, where the Shallow Auxiliary Fusion (SAF) and Deep Auxiliary Fusion (AAF) modules, combined with a re-parameterized heterogeneous multi-scale module (RepHMS) and a global heterogeneous flexible kernel selection mechanism (GHFKS), effectively improve feature fusion and expand the receptive field. This design significantly enhances multi-scale object detection performance while maintaining accuracy. Jiang et al. [29] introduced a new Multi-Branch Convolutional Neural Network (MBCNN), which uses three parallel convolution branches with different kernel sizes to expand the receptive field. This enables effective extraction of local features and dependencies across multiple receptive fields, which, when combined with higher-level features, achieves robust and high-accuracy recognition. Li et al. [30] proposed a Multi-Branch Attention Network (MBAN) specifically for small object detection. The network incorporates a Multi-Branch Attention Module (MBAM) with a dual-path structure, where features are processed in parallel using different convolution kernels. This approach preserves fine details of small objects while enabling cross-channel information interaction via 1 × 1 convolutions. Subsequently, the SimAM attention mechanism enhances critical features, resulting in more detailed small object representations and strengthened feature content.

2.3. Attention-Based Cross-Layer Feature Enhancement

To improve detection precision for small objects, beyond employing multi-branch convolutional pathways for feature extraction, an effective strategy involves applying attention mechanisms to refine the feature maps generated by the neck module. This strategy improves detection performance by integrating local and global information.

Du et al. [31] proposed a Cross-Layer Feature Pyramid Network (CFPT) for detecting small objects in aerial imagery, that incorporates a Cross-Layer Channel Attention (CCA) module and a Cross-Layer Spatial Attention (CSA) module. In CCA, feature maps are grouped by channel to perform cross-attention across the feature hierarchy, allowing each channel to access spatial context information from other channels. In CSA, each feature map is divided into multiple windows, and key-value interactions are performed between corresponding windows across layers, achieving cross-layer spatial information fusion. This design enhances both the semantic representation and detection sensitivity of small objects. Wang et al. [16] presented an efficient feature fusion framework to address the imbalance between classification and localization information [32]. The framework consists of a Fusion Downsampling (FDS) module, a Fusion Upsampling (FUS) module, and a Fusion Multi-Head Self-Attention (FMSA) module. The FDS module leverages local attention to strengthen the extraction and retention of fine details, the FUS module introduces global attention to preserve richer semantic information, and the FMSA module fuses multi-scale features from different layers via global multi-head self-attention. This framework effectively improves both detail extraction and semantic processing for small object detection, resulting in notable accuracy gains. Wang et al. [33] introduced an innovative Multi-Scale Global Detail Feature Integration Strategy (MGDFIS), which incorporates the FusionLock-TSS attention module, a global-to-detail fusion module, and a dynamic pixel attention module. This strategy strengthens the extraction of fine object details and improves the representation of complex features, providing an efficient solution to the limitations in feature expressiveness and imbalanced information flow in detecting small-scale objects. Hang et al. [34] addressed issues of complex background interference and low detection accuracy in UAV small object detection with the PARE-YOLO algorithm. In the neck, an Adaptive Temporal Feature Integration Module and an RT-DETR detection head based on Transformer perform attention-weighted fusion across feature maps of varying scales. This strategy markedly enhances both the reliability and accuracy of small-object detection by UAV under complex and adverse conditions.

3. Methodologies

In UAV images, small objects use few pixels, often resulting in inadequate feature information and insufficient multi-scale feature fusion. In order to overcome these difficulties, we design a feature enhancement network for small object detection, termed DMF-Net, based on multi-branch convolutional structures and attention mechanisms. DMF-Net mainly consists of two key modules: MBDE, which strengthens feature representation and contextual information integration, and DAF, which adaptively reweights adjacent-scale features to optimise the efficacy of multi-scale feature fusion.

3.1. Overview

As a high-performance one-stage object detection model, YOLO11 demonstrates excellent performance for general object detection tasks. However, in UAV small object detection scenarios, where targets are small and occupy few pixels, the progressive downsampling and feature pyramid structures used by YOLO11 tend to weaken fine-grained features in deep network layers, making it difficult to fully preserve small object representations. Moreover, its feature fusion process primarily relies on predefined multi-scale information paths, which limits the exploitation of complementary advantages between semantic and detailed information across different scales, thereby reducing the model’s capacity to represent and discriminate small objects in complex backgrounds. Specifically, in UAV small object detection, YOLO11n still faces three limitations. First, the conventional backbone design lacks a dedicated structure to preserve stable semantic information while progressively aggregating fine-grained local details. Second, its serial downsampling and deep convolution operations tend to progressively weaken local texture cues, which are critical for representing tiny objects with low pixel occupancy. Third, its neck mainly performs cross-scale interaction through predefined fusion paths, which is less flexible in modeling the adaptive complementarity between high-level semantic features and low-level detailed features in complex backgrounds.

Using YOLO11n as the baseline, this paper introduces a model named DMF-Net for small object detection. As depicted in Figure 2, the method consists of three main components: Backbone, Neck, and Head. In the backbone, the MBDE module is introduced to enhance feature representation by preserving stable semantic information and progressively aggregating fine-grained local details. Additionally, the MSDC-Conv module employs parallel convolutional paths to model local details across different scales and directions, thereby alleviating the weakening of local texture cues during deep feature extraction and significantly improving the network’s feature extraction capability. During the feature fusion stage, the DAF module is applied, which adaptively reweights adjacent-scale features using an attention mechanism, thereby improving the flexibility of cross-scale interaction and enhancing the complementarity between high-level semantic features and low-level detailed features for small object representation. In the detection stage, four enhanced feature maps that correspond to different scales are input to the detection head for classification and localization. Via simultaneously minimizing classification and regression losses across multiple scales, the model produces more accurate detection results.

3.2. DAF

In the domain of object detection, the identification of small objects poses significant challenges due to variations in scale, strong background interference and poor target-to-background contrast. To address these issues, mainstream detection frameworks typically employ multi-scale feature pyramids to fuse features from different levels. However, existing feature fusion methods largely rely on fixed top-down or bottom-up information paths, which limits their ability to fully capture the complementary relationship between high-level semantic features and low-level detailed features. Consequently, critical information for small objects can still be weakened in complex scenes. Furthermore, the contribution of features that are present at varying scales is not constant in small object detection, and fixed fusion strategies lack adaptive modeling of feature importance. In order to overcome the shortcomings, we design the DAF, as depicted in the lower right of Figure 2. DAF adaptively reweights adjacent-scale features and enables bidirectional information interaction, effectively enhancing the preservation and propagation of information pertinent to small objects throughout the multi-scale feature integration process. Specifically, DAF is the recipient of high-level semantic features and low-level detailed features from adjacent scales, which are provided as inputs. After channel alignment, adaptive attention weights are generated for each feature path, guiding bidirectional interaction and fusion. Unlike conventional bidirectional fusion, which mainly emphasizes top-down and bottom-up information propagation, the proposed DAF performs symmetric fusion at the fusion-unit level, where adjacent-scale features are assigned independent attention weights and updated through structurally mirrored operations. Specifically, the fusion process follows a symmetric design, in which high-level semantic features and low-level detail features are processed through structurally mirrored operations, enabling balanced and reciprocal information exchange across adjacent scales. This process simultaneously strengthens semantic and detailed information in the fused features, providing the detection head with more effective multi-scale representations.

It should be noted that the term bidirectional fusion in conventional feature pyramid designs usually refers to the existence of both top-down and bottom-up information propagation paths. Such designs improve cross-scale information flow, but their fusion operations are often still based on predefined routes and simple aggregation, without explicitly modeling balanced mutual enhancement between adjacent-scale features. By contrast, the proposed DAF performs symmetric fusion at the fusion-unit level. Specifically, the high-level semantic feature and the low-level detail feature are first aligned in channel space, then assigned independent attention weights, and finally updated through structurally mirrored operations. Therefore, the proposed symmetric fusion is not merely a bidirectional path design, but a balanced dual-branch interaction mechanism in which both adjacent-scale features serve as both information providers and receivers during fusion.

As illustrated in Figure 2, the DAF modules in the neck are arranged serially along the bidirectional fusion pathway rather than being used as isolated fusion units. Specifically, three DAF modules are connected sequentially in the top-down path and another three DAF modules are connected sequentially in the bottom-up path. In the top-down process, the high-level semantic feature serves as the main propagated stream, while the adjacent low-level detailed feature is introduced at each stage as complementary information to refine and enhance the current representation. In the bottom-up process, the fused low-level feature becomes the main propagated stream and further interacts with the adjacent high-level representation from the previous fusion stage. Therefore, the output of each DAF is forwarded to the next adjacent-scale DAF stage, forming a progressive cross-scale interaction process throughout the neck.

Denote the feature maps from two adjacent scales as ${\mathrm{F}}_{\mathrm{H}}\in {\mathbb{R}}^{{\mathrm{C}}_{\mathrm{H}}\times {\mathrm{H}}_{\mathrm{H}}\times {\mathrm{W}}_{\mathrm{H}}}$ and ${\mathrm{F}}_{\mathrm{L}}\in {\mathbb{R}}^{{\mathrm{C}}_{\mathrm{L}}\times {\mathrm{H}}_{\mathrm{L}}\times {\mathrm{W}}_{\mathrm{L}}}$. Here, ${\mathrm{F}}_{\mathrm{H}}$ represents the high-level feature map, which is abundant in semantic information, while ${\mathrm{F}}_{\mathrm{L}}$ represents the low-level feature map, containing abundant local detail information.

First, a 1 × 1 CBS (Conv–BN–SiLU) is utilized to both feature maps to perform channel mapping, aligning them to the same feature dimension:

$${\widehat{F}}_H={\varphi }_H\left(F_H\right),{\widehat{F}}_L={\varphi }_L\left(F_L\right),$$

(1)

Here, ${\varphi }_H$ and ${\varphi }_L$ denote CBS modules—composed of convolution, batch normalization, and a nonlinear activation function—used to unify the feature dimensions while preserving nonlinear representation capability.

Next, to adaptively capture the importance of features at different scales, DAF generates attention weights for both the high-level and low-level feature maps:

$$G_H=\sigma \left({\widehat{F}}_H\right),G_L=\sigma \left({\widehat{F}}_L\right),$$

(2)

Here, $\sigma \left(\right)$ denotes the Sigmoid function, which places a restriction on the attention weights, limiting them to the interval [0, 1].

Under the guidance of the attention weights, DAF performs bidirectional information interaction to simultaneously enhance both semantic and detailed features. This bidirectional interaction forms a symmetric feature enhancement process, where each scale contributes complementary information to the other in a balanced manner. The high-level feature map, enhanced by the low-level semantic information, is updated as:

$${\tilde{F}}_H={\widehat{F}}_H+{\widehat{F}}_H\odot G_H+(1-G_H)\odot \mu ({\widehat{F}}_L\odot G_L\to {\widehat{F}}_H),$$

(3)

Similarly, the low-level feature map, enhanced by high-level detailed features, is updated as:

$${\tilde{F}}_L={\widehat{F}}_L+{\widehat{F}}_L\odot G_L+(1-G_L)\odot \mu ({\widehat{F}}_H\odot G_H\to {\widehat{F}}_L),$$

(4)

Here, ⊙ denotes element-wise multiplication, and $\mu (·\to ·)$ denotes a scale alignment operator that transforms the source feature to the spatial size of the target feature before cross-scale interaction. The low-level feature ${\widehat{F}}_L$ is first enhanced by its corresponding attention weight $G_L$, and the term ${\widehat{F}}_L\odot G_L$ emphasizes informative responses within the low-level branch itself. Similarly, the attention-weighted high-level feature ${\widehat{F}}_H\odot G_H$ is aligned to the spatial size of the low-level branch through $\mu (·\to ·)$, enabling element-wise interaction across adjacent scales. The term $(1-G_L)$ acts as a complementary gate, allowing the low-level branch to selectively absorb additional semantic information from the high-level branch in regions where its own response is relatively weak.

Finally, these aligned high-level and low-level features get concatenated and fused:

$$F_{out}={\mathrm{C}}_{3\times 3}\left(\mathrm{Concat}({\tilde{F}}_H,{\tilde{F}}_L)\right),$$

(5)

Here, $C_{k\times k}$ denotes a k×k convolution composed of a CBS block.

The proposed DAF introduces a bidirectional attention-based feature interaction mechanism, establishing a more flexible and adaptive information propagation path between adjacent scales. This allows high-level semantic features and low-level detailed features to complement and reinforce each other during fusion. Compared with conventional multi-scale structures that rely on fixed fusion strategies, DAF effectively mitigates the feature representation deficiencies of small objects amid scale variations and cluttered scenes, providing the detection head with better distinguishable and reliable feature representations.

3.3. MBDE Design

In UAV object detection tasks, targets typically exhibit small scales, low pixel occupancy, and complex backgrounds, placing higher demands on the capacity of the backbone network to extract features. On one hand, the network must efficiently capture detailed features such as local edges and textures to support precise localization; on the other hand, it needs to extract stable high-level semantic information to ensure reliable class discrimination. However, conventional convolutional blocks generally process all channels uniformly, which can mix semantic information with local details, leading to localization errors and object confusion in densely populated or complex scenes. Additionally, consecutive downsampling and limited receptive fields cause high-frequency details to gradually attenuate, making it difficult to preserve small objects’ edges and structural information. To address these challenges, The backbone network incorporates MBDE and MSDC-Conv modules. As depicted in Figure 3, several serial MBDE and MSDC-Conv modules are incorporated to enhance local detail extraction while retaining original semantic information. The MBDE module introduces a multi-branch parallel structure within the backbone, representing the same feature layer across different receptive fields and feature subspaces, thereby effectively enhancing both semantic and local detail features. The MSDC-Conv, serving as a fundamental building block, employs multiple parallel paths with different convolutional combinations to extract fine-grained features. It fuses multi-level information while maintaining spatial consistency, providing the backbone network with richer and more stable feature representations.

3.3.1. MBDE

As depicted in Figure 4, we propose the MBDE module, which introduces an explicit channel allocation and fusion mechanism at the module level. The input feature map first passes through a 1 × 1 CBS to perform channel reorganization, producing the feature map Z. Next, a split operation divides Z along the channel dimension into a bypass branch $Z_b$ and an enhancement branch $Z_e$. The bypass branch $Z_b$ directly preserves the original features, ensuring continuous propagation of class-relevant information. The enhancement branch $Z_e$ is refined progressively by a series of N MSDC-Conv blocks to strengthen the response to local details, with the enhanced features denoted as $X_i$, where $\left\{X_i\right\}(i\in \{1,\dots ,N\})$. Finally, the module output is obtained by concatenating the original feature $Z_b$ with the N enhanced features $X_i$ across the channel dimension. This design facilitates the simultaneous retention of stable semantic information and locatable local detail features by the output, thereby enhancing the model’s capacity for small object detection under complex conditions.

The MBDE adopts a structured design of channel splitting, progressive enhancement, and multi-level aggregation to effectively organize the feature extraction process. On one hand, the bypass branch preserves stable original information, reducing interference during semantic propagation and improving training stability. On the other hand, the enhancement branch stacks N MSDC-Conv blocks to progressively refine local structural features, with outputs from each stage aggregated to the final output. This design ensures that the output features simultaneously contain multi-depth local detail information and semantic information. With minimal additional computational overhead, MBDE improves feature separability and localization consistency, rendering it especially suitable for UAV object detection in cluttered scenes.

3.3.2. MSDC-Conv

Figure 5 shows the proposed MSDC-Conv adopts a multi-branch parallel structure for feature extraction from the input image. The input is fed into multiple independent convolutional paths, each consisting of serial convolutions with different combinations, designed to capture complementary local details and contextual information across various receptive fields and feature subspaces. Some branches perform feature enhancement via element-wise addition, preserving original details and stabilizing information propagation. Ultimately, the processed features of each branches are combined at the feature level to form a unified representation, which significantly enriches feature diversity and discriminability while maintaining spatial consistency.

The input image is herein designated $\mathbf{I}\in {\mathbb{R}}^{H\times W\times C}$, wherein C, H, and W represent the number of channels, height, and width, correspondingly. First, the input image is replicated four times and fed into the corresponding branch paths for information capture. The feature extraction process of the (i)-th branch can be formulated as:

$$F_i={\mathcal{F}}_i\left(I\right),i\in \{1,2,3,4\},$$

(6)

Here, $\mathcal{F}\left(\right)$ denotes a nonlinear mapping function composed of several serial convolution operations, corresponding to the convolutional paths in Figure 5 formed by 1 × 1 CBS, asymmetric convolutions, and 3 × 3 CBS combinations. By employing serial convolutions with different structures, each branch can extract local details and contextual information across various scales and orientations.

For the first two branches, the Local Residual Detail Branches (LRDBs), denoted as ${\mathcal{F}}_1$ and ${\mathcal{F}}_2$, the input image is first duplicated. The first copy sequentially passes through 1 × 1 CBS, 1 × 3 CBS, 3 × 1 CBS, and 3 × 3 CBS convolutions, enabling multi-level detail feature extraction. The second copy is then combined with the output of the first path, $F_i$, via a local residual connection to alleviate feature degradation in deep layers:

$${\tilde{F}}_i={\mathrm{C}}_{3\times 3}\left({\mathrm{C}}_{3\times 1}\left({\mathrm{C}}_{1\times 3}\left({\mathrm{C}}_{1\times 1}\left(I\right)\right)\right)\right),i\in \{1,2\},$$

(7)

$$F_i={\tilde{F}}_i+I,$$

(8)

For the third branch, the Compact Semantic Branch (CSB), denoted as ${\mathcal{F}}_3$ in Figure 2, a combination of 1 × 1 CBS, 3 × 3 CBS, and 1 × 1 CBS is employed. This design strengthens high-level semantic representation while keeping computational overhead under control, formulated as:

$$F_3={\mathrm{C}}_{1\times 1}\left({\mathrm{C}}_{3\times 3}\left({\mathrm{C}}_{1\times 1}\left(I\right)\right)\right),$$

(9)

For the fourth branch, the Identity Mapping Branch (IMB), denoted as ${\mathcal{F}}_4$ in Figure 2, a standalone 1 × 1 CBS aims to keep the basic representation of the initial features, providing stable complementary information, expressed as:

$$F_4={\mathrm{C}}_{1\times 1}\left(I\right),$$

(10)

Ultimately, those features of all the constituent paths are aggregated across the channel axis. Here’s how the module’s result is expressed:

$$F_{fusion}=\mathcal{G}\left(\mathrm{Concat}(F_1,F_2,F_3,F_4)\right),$$

(11)

Here, $\mathrm{Concat}\left(\right)$ denotes concatenation along the dimension of the channel, and $\mathcal{G}\left(\right)$ represents a nonlinear activation function. This fusion strategy allows feature responses from different receptive fields and convolutional paths to complement each other at the same spatial locations, thereby simultaneously preserving the local details of small objects and the necessary contextual information.

By introducing a multi-branch serial convolution structure combined with an effective fusion mechanism, MSDC-Conv is able to preserve the original feature information while keeping computational complexity and parameter size under control. At the same time, it enables the capture of fine-grained features across multiple directions as well as scales, significantly enhancing the detail representation and discriminative capability of small UAV targets. This provides the reliable foundation required for fusion and stable predictions.

4. Experiments

4.1. Datasets

In order to test the designed method’s capability in detecting UAVs in complex backgrounds, this study employs two publicly available and widely used UAV datasets: VisDrone (Vision Meets Drones) and TinyPerson (Tiny Person Detection Dataset).

The VisDrone dataset is a prominent example of a frequently utilised dataset for the domain of UAV object detection. It contains images captured by various types of drones across 14 cities in China, covering complex scenarios such as squares, streets, campuses, and commercial areas, offering high scene diversity. Because of variations in drone fly height, camera angles, and lighting conditions, the dataset also exhibits diverse viewing perspectives. VisDrone annotates a total of ten object categories: bicycle, awning-tricycle, tricycle, van, bus, truck, car, pedestrian, people, and motor. The image resolution is approximately 2000 × 1500 pixels. As shown in the left image of Figure 6, the dataset includes urban streets, highways, nighttime scenes, and suburban areas, featuring rich visual content and strong background interference. The UAV acquisition process introduces significant target size and shape distortion. Objects vary drastically in scale, with a high proportion of small, distant instances. Lighting conditions fluctuate considerably, and nighttime images contain noise and glare. Targets are often densely distributed, leading to occlusion and overlap, which further complicates accurate localization and classification.

TinyPerson is a dataset with high spatial resolution, specifically constructed to support pedestrian detection at very small scales in complex scenes. It is primarily used to evaluate a model’s localization and recognition capability for extremely small pedestrian targets. The dataset is collected from various high-altitude imaging platforms, mainly covering open outdoor environments such as beaches and coastlines. As illustrated in the image on the right of Figure 6, TinyPerson focuses on pedestrian detection under long-distance, top-down viewpoints. The pedestrian instances are extremely small and densely distributed, while frequently suffering from heavy occlusion and background clutter. Due to the prevalence of small targets and the wide variation in object scales, this dataset presents significant challenges in terms of effective feature extraction and robust multi-scale representation learning. According to the official TinyPerson benchmark [35], most target sizes are below 20 pixels. In addition, as shown in Figure 7b, the normalized widths and heights of most objects are concentrated below 0.1, further quantifying the extremely small scale of the dataset.

As shown in Figure 7, we analyzed and plotted the normalized width–height distributions of all annotated objects in the TinyPerson and VisDrone datasets. Widths and heights of all objects are normalized by dividing by the corresponding image width and height, respectively, so that the values range from 0 to 1. Accordingly, a normalized value of 0.1 corresponds approximately to 10% of the image width or height, and 0.2 corresponds to 20%. This normalization enables comparison of object scales across images with different resolutions. In Figure 7a, the VisDrone dataset exhibits a widely dispersed object scale distribution, covering numerous small and medium-sized targets. This reflects the multi-scale nature of real-world UAV surveillance scenarios, increasing the difficulty of feature alignment and discrimination for objects of different sizes. In contrast, Figure 7b shows that the TinyPerson dataset primarily contains extremely small-scale targets, with most objects having normalized widths and heights below 0.1 and being densely distributed. This serves to underscore the difficulties inherent in the detection of small objects. This dataset has a limited range of object scales, with few large targets, requiring models to possess strong high-resolution feature extraction and fine-grained localization capabilities.

Figure 8 presents the per-category instance count in the VisDrone (a) and TinyPerson (b) datasets used in this study, both exhibiting clear class imbalance. VisDrone contains 10 object categories, with car and pedestrian dominating at 187,005 and 109,187 instances, respectively, while categories such as awning-tricycle, tricycle, and bus have relatively few samples, reflecting a typical long-tail distribution that increases detection difficulty for underrepresented and easily confused classes. TinyPerson contains only two categories, earth_person and sea_person, with earth_person significantly more prevalent. The long-tail effects and cross-scene imbalances in these datasets closely mirror real-world applications, providing a rigorous benchmark for evaluating model robustness under complex conditions.

4.2. Experimental Setup

All models were evaluated under the same environment and all baseline models used identical hyperparameters. The experimental setup is outlined below: the operating system is Red Hat 4.8.5-28; programming is conducted in Python 3.11; the deep learning framework is PyTorch 2.1; CUDA 12.1 is used for GPU acceleration; the GPU is an NVIDIA A800 80 GB with 80 GB of VRAM; and the CPU is an Intel Xeon 6338N. The experimental hyperparameters are summarized in

Table 1. Model training hyperparameters.

Feature	Parameter
Optimizer	SGD
Epochs	300
Batch size	32
Input image-size	640 × 640
base_learning_rate	0.01
weight_decay	0.0005
Momentum	0.937
Workers	4
Mixed precision	✓
Learning rate schedule	Cosine

. For training on both VisDrone and TinyPerson datasets, standard data augmentation strategies were applied to improve model generalization and performance on small, dense targets. Specifically, we applied horizontal flipping, color jittering, mosaic augmentation, and random cropping as data augmentation techniques. These augmentation techniques ensure robust feature learning and support reproducibility.

4.3. Evaluation Metrics

In the present experiments, the following evaluation metrics are employed: Precision, Recall, mAP_50, and mAP_50:95. Precision is defined as the proportion of correctly predicted results out of all predictions made, reflecting the accuracy of the model’s detections and its reliability in producing correct results. High precision indicates that the model can effectively suppress background interference while correctly identifying targets, even under complex conditions.

$$\mathrm{precision}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}},$$

(12)

Recall reflects the fraction of ground-truth objects that are successfully identified by the model, indicating its effectiveness in detecting all targets and reducing missed detections. A high recall indicates that the model can accurately detect real targets under complex backgrounds, reducing the frequency of missed detections and improving coverage of scene objects.

$$\mathrm{recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}},$$

(13)

Since Precision and Recall evaluate only false positives and false negatives individually, they cannot provide a comprehensive assessment of a model’s overall detection capability. Therefore, we adopt mAP_50 as an integrated evaluation metric. mAP_50 measures the model’s ability to detect true objects at an IoU threshold of 0.5, calculated by integrating the Precision–Recall curve to obtain the average precision (AP) for each class, and then averaging over all object categories.

$$\mathrm{mAP}={\displaystyle \frac{{\sum }_{n=1}^N{\mathrm{AP}}_n}{N}},$$

(14)

In the above formulation, AP is derived from the integral of the Precision–Recall curve, and larger AP values imply stronger detection performance for the associated category. ${\mathrm{AP}}_n$ denotes the average precision for the n-th object class, expressed as:

$$\mathrm{AP}={\int }_0^1\mathrm{Precision}\left(\mathrm{R}\right)\mathrm{dR},$$

(15)

In the above formulation, P(R) represents the precision at different recall values.

mAP_50 primarily evaluates whether a model correctly detects true objects, but its single IoU threshold of 0.5 is relatively lenient and cannot fully reflect detection accuracy under stricter localization requirements. To address this, we adopt mAP_50:95 for a more comprehensive assessment. In this metric, the IoU threshold starts at 0.5 and is incremented by 0.05 up to 0.95. For each class, AP is computed at all IoU thresholds, then averaged across all thresholds. Finally, the mean over all object classes is calculated to obtain mAP_50:95, providing a more complete evaluation of the model’s detection performance across varying localization strictness.

$$\mathrm{mAP}\_50:95={\displaystyle \frac{1}{C}}\sum _{c=1}^C\left({\displaystyle \frac{1}{10}}\sum _{t=0.5}^{0.95}{\int }_0^1{\mathrm{P}}_c^{\left(\mathrm{t}\right)}\left(\mathrm{R}\right)\mathrm{dR}\right),$$

(16)

In the above formula, c denotes the class index, and t represents the IoU threshold with $t\in \{0.5,0.55,\dots ,0.95\}$. ${\mathrm{P}}_c^{\left(\mathrm{t}\right)}$ is the average precision for class c at the IoU threshold t.

In addition to the aforementioned evaluation metrics, we also report GFLOPs, Params, and FPS to provide a supplementary assessment of the models. GFLOPs reflects the model’s computational burden and overall efficiency, Params indicates the model size and deployment overhead, and FPS measures the inference speed and real-time performance. Together, these metrics allow for a more comprehensive comparison of different methods with respect to both accuracy and efficiency.

4.4. Ablation Experiments

This study used YOLO11n as the baseline model in ablation experiments on the VisDrone and TinyPerson datasets. Under consistent experimental settings and hyperparameters, both single-module ablation and full-model ablation studies were performed.

4.4.1. Single-Module Ablation

In order to accurately evaluate the role of the presented MSDC-Conv downsampling module at different stages within the YOLO11n framework, we adopt a layer-wise cumulative single-point ablation strategy, as depicted in Figure 9. This design uses the original YOLO11n as the baseline, while preserving other parts of the network and training configuration unchanged, gradually introduces the MSDC-Conv module from shallow to deep layers. Figure 9a corresponds to inserting MSDC-Conv at the p1 downsampling layer, while the remaining downsampling layers retain the original convolutional structure. Building on this, Figure 9b shows the further addition of MSDC-Conv at the p2 layer; subsequently, Figure 9c and Figure 9d depict the cumulative replacement at p3 and p4 layers, respectively. Ultimately, as depicted in Figure 9e, MSDC-Conv is applied to all downsampling layers from p1 to p5, forming the complete enhanced model. Unlike a one-shot full replacement, this layer-wise cumulative ablation design provides a clearer assessment of each downsampling stage’s independent contribution to detection performance. On one hand, it effectively avoids performance coupling interference caused by simultaneous modifications across multiple layers; on the other hand, it facilitates analysis of MSDC-Conv’s distinct roles in preserving fine-grained shallow features and modeling deep semantic representations.

Under the progressive layer-wise single-point ablation setting illustrated in Figure 9, the downsampling convolutions in the backbone network were gradually replaced with MSDC-Conv on the VisDrone dataset. As reported in

Table 2. Single-point Ablation Experiment on VisDrone Dataset.

Model	Params (M)	GFLOPs	Precision	Recall	mAP_50	mAP_50:95
YOLO11n	2.5	6.3	46.3	33.0	33.7	19.6
(a)	2.5	7.9	46.0	35.1	35.2	20.0
(b)	2.6	9.4	46.9	36.9	36.9	22.0
(c)	2.6	10.8	47.0	37.7	37.8	22.6
(d)	2.8	12.2	48.2	38.1	38.2	23.0
(e)	3.9	13.5	49.0	37.6	38.2	22.9

Note: Bold values indicate the best performance for the corresponding metric.

, the performance exhibits a clear pattern of cumulative improvement. The baseline YOLO11n achieves mAP_50 and mAP_50:95 scores of 33.7% and 19.6%, respectively. When MSDC-Conv is introduced at the P1 downsampling layer, mAP_50 and mAP_50:95 increase to 35.2% and 20.0%, indicating that improved downsampling plays a direct role in preserving target details and enhancing localization stability in aerial imagery. As the replacement is extended toward the middle and deeper layers, detection performance continues to improve. Among all configurations, setting (d) achieves the best mAP_50:95 and Recall, while setting (e) attains the same mAP_50 with higher Precision but also higher computational cost. This suggests that for datasets such as VisDrone, which are characterized by large scale variations and complex background textures, improving downsampling at middle-to-high levels is particularly critical for enhancing the overall discriminative capability. Notably, when MSDC-Conv is further applied to all downsampling layers (configuration (e)), Precision continues to increase, while Recall and mAP_50:95 show a slight decline. This phenomenon indicates that excessive replacement may lead to performance saturation or minor performance fluctuations. Overall, these results demonstrate that MSDC-Conv effectively improves detection performance on VisDrone; however, practical configuration choices should carefully balance accuracy gains against additional computational overhead. On TinyPerson (

Table 3. Single-point ablation experiments on the TinyPerson dataset.

Model	Params (M)	GFLOPs	Precision	Recall	mAP_50	mAP_50:95
YOLO11n	2.5	6.3	39.6	27.6	23.9	7.6
(a)	2.5	7.9	40.8	27.8	24.9	7.8
(b)	2.6	9.4	39.4	29.4	25.3	8.2
(c)	2.6	10.8	38.9	28.6	24.9	8.1
(d)	2.8	12.2	41.5	28.4	24.8	7.7
(e)	3.9	13.5	42.6	30.6	25.8	8.1

Note: Bold values indicate the best performance for the corresponding metric.

), configuration (e) achieves the best overall detection performance. Therefore, configuration (e) is adopted as the default setting based on a comprehensive consideration of dataset characteristics, overall performance, and implementation consistency.

In contrast to VisDrone, TinyPerson places greater emphasis on the detectability of extremely small targets. Table 3 reports the results of applying the same layer-wise replacement strategy on the TinyPerson dataset. The baseline YOLO11n achieves mAP_50 and mAP_50:95 of 23.9% and 7.6%, respectively. When MSDC-Conv is introduced only at the shallow layers, mAP_50 and mAP_50:95 increase to 24.9% and 7.8%. Further extending the replacement to configuration (b) leads to an improvement in mAP_50:95 to 8.2%, while Recall also rises to 29.4%. As the replacement is progressively extended to deeper layers, the performance gains of configurations (c) and (d) are no longer strictly monotonic, indicating a more complex coupling between architectural modifications and performance improvements in the TinyPerson scenario. Ultimately, the full-layer replacement configuration (e) achieves the best performance, with the highest mAP_50, Precision, and Recall of 25.8%, 42.6%, and 30.6%, respectively, while maintaining a relatively high mAP_50:95.

Based on the single-point ablation results on both VisDrone and TinyPerson, several conclusions can be drawn. On VisDrone, configuration (e) achieves the same mAP_50 as configuration (d) while further improving Precision to 49%, indicating that (e) is appropriate for detection of small objects. On TinyPerson, the full replacement strategy (e) yields higher detection capability, which is particularly beneficial for extremely small targets. Considering that this work focuses on stable detection of tiny objects under complex backgrounds, and aims to adopt a unified and reproducible architecture to facilitate subsequent training and deployment, configuration (e) is ultimately selected as the default model setting. Specifically, MSDC-Conv is introduced at all downsampling stages from p1 to p5. This configuration provides consistent performance gains on both datasets and demonstrates more coherent advantages in cross-scene generalization and sensitivity to small objects. Therefore, the remaining configurations are not adopted as the final solution.

4.4.2. Overall Ablation

We conducted an ablation study on the VisDrone dataset to evaluate the effect of each innovative module on UAV small object detection. The modules focused on were MSDC-Conv, MBDE and DAF, as shown in

Table 4. Ablation Study on the VisDrone Dataset.

MSDC-Conv	MBDE	DAF	Params (M)	GFLOPs	Precision	Recall	mAP_50	mAP_50:95
×	×	×	2.5	6.3	46.3	33.0	33.7	19.6
✓	×	×	3.9	13.5	49.0	37.6	38.2	22.9
✓	✓	×	4.4	16.9	49.1	39.9	39.7	23.9
✓	✓	✓	5.5	26.6	52.9	42.3	44.0	27.2

Note: Bold values indicate the best performance for the corresponding metric.

. When no modules were introduced into the baseline model, the mAP_50 was 33.7%, Precision was 46.3%, and Recall was 33.0%. After incorporating the MSDC-Conv module, the model’s capacity to extract multi-scale detail information was strengthened, with mAP_50 rising to 38.2%, mAP_50:95 increasing to 22.9%, and Precision and Recall improving to 49.0% and 37.6%, respectively. The further introduction of the MBDE module strengthened the multi-scale modeling capability, resulting in a mAP_50 increase to 39.7% and Recall rising to 39.9%, indicating the contribution of MBDE in detecting small objects in complex backgrounds. When all three modules were used together, the DAF module effectively suppressed background noise and enhanced key feature representation through recalibration of the feature pyramid and context fusion. This led to a mAP_50 of 44.0%, an mAP_50:95 increase to 27.2%, and Precision and Recall reaching 52.9% and 42.3%, respectively. The combined effect of these modules has been found to significantly enhance the accuracy with which small objects are detected, as well as the capacity to effectively integrate features across multiple scales. Although the introduction of MBDE and DAF increases Params and GFLOPs, the resulting gains in Precision, Recall, mAP_50, and mAP_50:95 are substantial, indicating that the additional computational cost is acceptable for UAV small object detection scenarios that prioritize detection accuracy and robustness.

Table 5. Ablation Study on the TinyPerson Dataset.

MSDC-Conv	MBDE	DAF	Params (M)	GFLOPs	Precision	Recall	mAP_50	mAP_50:95
×	×	×	2.5	6.3	39.6	27.6	23.9	7.6
✓	×	×	3.9	13.5	42.6	30.6	25.8	8.1
✓	✓	×	4.4	16.9	43.8	31.3	26.8	8.8
✓	✓	✓	5.5	26.6	46.0	34.7	30.3	9.7

Note: Bold values indicate the best performance for the corresponding metric.

presents the results of the ablation study on the TinyPerson dataset. Similar to the findings on the VisDrone dataset, as different modules were progressively combined on the baseline model, the Precision, Recall, mAP_50, and mAP_50:95 metrics consistently increased, reaching their maximum values when all three modules were used simultaneously. While MSDC-Conv enhanced feature extraction capabilities, MBDE enabled the model to capture local dependencies between objects, helping the network better understand contextual relationships. DAF further optimized the feature fusion process, allowing the model to focus more effectively on critical regions. The optimised model exhibited substantial enhancements in precision and recall for the detection of tiny objects, thereby validating the efficacy of multi-scale perception and feature recalibration in scenarios characterised by densely populated small objects. Similarly, on the TinyPerson dataset, the increase in model complexity is accompanied by clear improvements in all major evaluation metrics, suggesting that the additional Params and GFLOPs are justified for extremely small object detection under complex backgrounds.

It should be noted that adding all three modules increases the computational cost from 6.3 GFLOPs to 26.6 GFLOPs, while the gain in detection performance is substantial. Moreover, the model still achieves 150.4 FPS (

Table 6. Comparative Experiments on the VisDrone Dataset.

Method	AP	AP_50	AP_75	AP_s	AP_m	AP_l	Params (M)	GFLOPs	FPS
Two-stages:
Faster-RCNN	20.3	34.0	21.7	9.8	29.4	43.1	41.3	208.0	46.8
Cascade-RCNN	20.5	33.5	22.2	9.9	29.8	45.1	69.2	236.0	24.6
One-stages:
TOOD	21.3	34.5	23.0	10.3	30.2	41.4	32.0	199.0	25.7
ATSS	21.5	34.5	23.1	10.0	30.7	45.7	38.9	110.0	35.4
RetinaNet	17.8	29.3	18.7	6.6	26.4	43.0	36.5	210.0	54.4
RTMDet	18.2	31.1	21.2	7.6	28.6	44.4	4.8	8.0	52.1
YOLOX	15.5	28.2	15.4	7.7	21.2	28.7	5.0	7.5	88.0
YOLO11n *	19.7	35.0	20.0	10.4	29.4	37.2	2.5	6.3	177.0
YOLO12n *	13.6	24.9	13.3	5.4	21.3	32.4	2.5	5.8	204.7
YOLO13n *	13.4	24.5	13.2	5.5	21.2	30.6	2.5	6.1	177.3
RT-DETR *	19.8	31.2	7.0	10.3	30.1	39.0	38.6	57.0	147.9
DMF-Net *	19.8	35.5	21.9	10.6	29.6	37.3	5.5	26.6	150.4

Note: YOLO13n is based on a third-party implementation of YOLOv13-N (https://github.com/iMoonLab/yolov13) (accessed on 23 April 2026) rather than an official Ultralytics release. Methods marked with * were trained for 300 epochs in our own experimental environment, while the other baseline methods were trained for 12 epochs on the MMDetection platform.

) on our testing platform, meeting real-time inference requirements for many UAV applications. Therefore, future work will focus on lightweight designs, such as model pruning, quantization, and efficient module simplification, to further reduce computational overhead while maintaining high detection accuracy.

Figure 10 compares the feature map visualizations generated after the P5 layer for YOLO11n and DMF-Net on the TinyPerson dataset. Panel (a) shows the input scene. In the feature map of YOLO11n (Figure 10b), object responses are sparse and weak, the boundaries of densely packed targets are blurred, and features often mix with the background. In contrast, the feature map produced by the proposed MSDC-Conv (Figure 10c) exhibits continuous, high-intensity responses for dense small object instances, clearly separates the boundaries of closely packed targets to prevent instance confusion, and enhances contrast with low-distinction backgrounds. These improvements directly increase the feature discriminability of small objects in complex scenes, demonstrating that MSDC-Conv is better suited than standard YOLO11n convolution for small object detection tasks.

4.4.3. MBDE-Specific Feature Visualization Analysis

To provide more direct evidence for the role of MBDE, we further visualize the feature responses at the corresponding stage before and after inserting MBDE using samples from the TinyPerson dataset, as shown in Figure 11. Since TinyPerson contains extremely small and detail-sensitive targets, it is suitable for examining whether MBDE enhances fine-grained local responses. Compared with the baseline feature maps, the MBDE-enhanced features exhibit clearer activations around tiny-object regions and more distinct local structural responses, suggesting that MBDE helps preserve detail-sensitive information during feature extraction.

4.5. Comparison Experiments

A series of comparative experiments were conducted on the same dataset using state-of-the-art object detection algorithms. Specifically, baseline methods trained on the MMDetection platform were trained for 12 epochs, whereas select methods (marked with *) were trained for 300 epochs in our own experimental environment. In MMDetection, the training process employs an iteration-based scheduling strategy, pretrained weight initialization, and adaptive learning rate policies, which enable models to achieve sufficient parameter updates and reach stable convergence within a relatively small number of epochs. Therefore, the reduced number of epochs in MMDetection does not compromise the comparability of the results with those obtained from traditional long-term training. The experiments included representative two-stage detectors (Faster R-CNN, Cascade R-CNN) and one-stage detectors (TOOD, ATSS, RetinaNet, RTMDet, YOLOX, RT-DETR). Under identical hardware conditions, we consistently measured each model’s parameter count (Params), computational cost (GFLOPs), inference speed (FPS), and detection performance metrics (AP_50, AP_75, AP_50:95).

4.5.1. VisDrone Dataset

As shown in Table 6, two-stage detectors achieve relatively stable accuracy but incur high computational costs and limited speed. For instance, Cascade R-CNN has 69.2M parameters and 236 GFLOPs, with an FPS of only 24.6, making real-time deployment challenging. Faster R-CNN achieves a higher FPS of 46.8 but a modest AP_50 of 34.0%. In contrast, one-stage detectors are generally more suitable for real-time scenarios: TOOD and ATSS both reach an AP_50 of 34.5%, with ATSS slightly outperforming on AP_75 at 23.1%, while also requiring less computation and offering faster inference. RetinaNet achieves 54.4 FPS but with relatively lower accuracy. Among lightweight models, YOLO11n strikes a better balance between accuracy and speed, reaching an AP_50 of 35.0% and an FPS of 177.0. YOLO12n and YOLO13n further improve speed (FPS 204.7 and 177.3, respectively) but achieve only moderate AP_50. RT-DETR also offers high speed but a lower AP_50 of 31.2%. In comparison, our method achieves the highest AP_50 while maintaining 150.4 FPS and reaches 21.9% on AP_75, demonstrating superior overall performance.

It should be noted that, on the VisDrone dataset, the improvement of DMF-Net over YOLO11n in AP_50 is relatively limited, while the model complexity increases in terms of parameters and GFLOPs. However, the gain is more evident under the stricter AP_75 metric, where DMF-Net improves from 20.0% to 21.9%, suggesting better localization performance under stricter evaluation criteria. Moreover, despite the increased computational cost, DMF-Net still achieves 150.4 FPS, which remains sufficient for real-time inference in many UAV applications. Therefore, the proposed model is more suitable for scenarios that require improved detection quality while maintaining real-time performance, rather than extremely resource-constrained deployment settings.

As shown in Figure 12, DMF-Net maintains stable detection performance across various challenging scenarios. In low-contrast, long-distance small-object scenes, such as nighttime or low-light environments, the enhanced MSDC-Conv and MBDE modules significantly improve the model’s capacity to capture subtle weak features and capture local dependencies, enabling consistent responses to distant targets and reducing missed detections. In dense traffic and parking areas, the DAF module facilitates high-quality multi-scale feature fusion, improving small-object detection, clarifying boundaries between adjacent targets, and producing tighter bounding boxes. These improvements increase object separability and localization stability, allowing the model to better adapt to complex environments.

Figure 13 presents visual comparison results for three representative examples from the VisDrone test set: (a) input images, (b) outputs from the baseline YOLO11n, and (c) outputs from DMF-Net. Visualisations use green for correct, red for missed and blue for false positives. Under challenging conditions commonly encountered in aerial urban scenes—such as distant small objects, highly clustered objects, occlusion, and strong background texture interference—the baseline model tends to show an increase in both red and blue boxes. Specifically, it produces false positives (blue boxes) in complex background regions such as building shadows and road edges, while distant or occluded small vehicles are often missed (red boxes). In contrast, DMF-Net exhibits a higher proportion of green boxes on the same examples, demonstrating more stable separation of dense targets, more complete detection of distant small objects, and suppression of irrelevant background activations. This results in improved robustness and generalization, effectively reducing both missed and false detections.

4.5.2. TinyPerson Dataset

Table 7. Comparative Experiments on the TinyPerson Dataset.

Method	AP	AP_50	AP_75	AP_s	AP_m	Params (M)	FPS
Two-stages:
Faster-RCNN	4.1	12.1	1.7	3.6	36.8	41.3	59.7
Cascade-RCNN	4.4	12.6	2.0	3.8	37.4	69.2	39.6
One-stages:
ATSS	4.0	13.8	1.1	3.7	21.7	38.9	46.6
RetinaNet	1.5	5.6	0.5	1.3	16.5	36.5	46.3
YOLOX	5.5	21.3	1.5	6.0	12.5	5.0	99.1
YOLO11n *	6.1	18.8	2.0	5.5	38.3	2.5	83.2
YOLO12n *	5.0	15.9	2.2	4.7	33.7	2.5	153.3
YOLO13n *	4.7	15.0	2.0	4.4	31.8	2.4	146.0
RT-DETR *	6.8	21.7	2.9	6.5	7.3	38.5	133.8
DMF-Net *	7.5	22.9	2.6	7.3	34.4	5.5	97.2

Note: Methods marked with * were trained for 300 epochs in our own experimental environment, while the other baseline methods were trained for 12 epochs on the MMDetection platform.

presents the comparison of mainstream detectors on the TinyPerson dataset. Overall, two-stage methods show limited advantages in this extreme small-object scenario: Faster R-CNN and Cascade R-CNN achieve APs of only 4.1% and 4.4%, respectively, while also suffering from large model size and limited inference speed. Among one-stage methods, RetinaNet and ATSS achieve APs of 1.5% and 4.0%, indicating insufficient adaptation to tiny pedestrian targets. Although YOLO series models offer high inference efficiency, their accuracy improvement is limited in TinyPerson, a dataset dominated by extremely small objects; for example, YOLOX and YOLO11n achieve APs of 5.5% and 6.1%, respectively. RT-DETR reaches 6.8% AP but still lags behind our method. In comparison, DMF-Net achieves 7.5% AP, 22.9% AP_50, and 2.6% AP_75, outperforming all other methods in overall accuracy. For small-object-specific metrics, AP_s reaches 7.3%, significantly higher than YOLO11n’s 5.5%, demonstrating that the proposed improvements more effectively enhance detection of extremely small-scale pedestrians. In terms of efficiency, our model contains 5.5M parameters while maintaining real-time inference at 97.2 FPS, realizing a superior balance between accuracy, model size, and speed. It should be noted that the AP of DMF-Net on the TinyPerson dataset is 7.5%, which appears low in absolute terms. However, this dataset primarily contains extremely small pedestrian targets with most object sizes below 20 pixels and densely distributed, presenting significant challenges for detection. In real-world UAV applications, even a relatively low AP can provide valuable detections for downstream tasks such as tracking, alerting, and scene understanding. Therefore, while the AP is low, the model still offers practical utility in challenging small-object detection scenarios.

Figure 14 shows several detection examples of DMF-Net on the TinyPerson dataset. Under these challenging conditions, DMF-Net consistently generates stable responses for numerous tiny pedestrians, with the overall bounding box distribution closely matching the actual crowd locations. In background regions with strong interference, such as the sea and shoreline, the model outputs are restrained, avoiding large-scale false activations and demonstrating effective suppression of complex background noise.

Figure 15 compares the detection visualizations of YOLO11n and DMF-Net on the TinyPerson dataset. As shown in panels (b) and (c), the baseline model produces false positives in regions with complex textures and varying illumination, while missing distant pedestrians (blue boxes). In contrast, DMF-Net generates more detection boxes that align with the true object locations (green boxes), with both missed detections (red boxes) and false positives (blue boxes) reduced. These results further demonstrate DMF-Net’s strong generalization ability and its sensitivity to tiny objects in complex backgrounds, achieving superior performance.

Figure 16 presents prediction heatmaps of various models on the TinyPerson dataset. Panel (a) displays the input image, panels (b)–(d) show the heatmaps for YOLO11n, YOLO12n, and YOLO13n, respectively, and panel (e) shows the results of DMF-Net. All methods use consistent input settings and post-processing, with predicted responses for the target class visualized as heatmaps. From the visualizations, YOLO11n exhibits unstable responses in small-object and low-contrast regions, appearing as fragmented or discontinuous hotspots, indicating that small-scale pedestrian representations are heavily affected by noise. YOLO12n shows responses closer to the main targets but only scattered activations in complex background areas, suggesting insufficient background suppression. YOLO13n produces stronger responses in dense regions, with hotspots expanding and potentially merging with adjacent targets, which can reduce object resolution and localization accuracy. In contrast, DMF-Net’s high-response regions align more closely with pedestrian locations, with concentrated hotspots that maintain clear instance separation, particularly in crowded or tiny-object areas, while effectively controlling background activations.

5. Conclusions

The present work tackles the difficulty of limited accuracy in UAV small-object detection under complex backgrounds by developing a detection framework centered on feature extraction and cross-scale fusion, and proposing a detection network incorporating MBDE and DAF modules. First, from the perspective of feature extraction, we analyze the limitations of deep serial convolutional structures in modeling high-frequency details and local textures for small object detection. In order to tackle this, we introduce MBDE, which incorporates MSDC-Conv to implement multi-branch parallel processing with different local receptive fields. Combined with residual connections to preserve original semantic information, this design significantly enhances fine-grained feature representation without compromising semantic consistency. Second, we propose the DAF module, which leverages an attention mechanism to adaptively weight features across adjacent scales and dynamically adjust cross-scale information interaction. The symmetric interaction structure between adjacent-scale features further improves the consistency and robustness of multi-scale representation learning. The feature fusion strategy based on DAF effectively balances local discriminative cues and high-level semantic information in complex and densely populated scenes, substantially improving the network’s overall perception of small objects.

Extensive comparative experiments show that the proposed method improves small-object detection performance on both VisDrone and TinyPerson datasets, with more evident gains on TinyPerson. Compared with the baseline, the proposed model improved AP_50 by 0.5% on the VisDrone dataset and by 4.1% on the TinyPerson dataset. Overall, the proposed method shows the potential to improve UAV small-object detection.

Notwithstanding the substantial enhancements accomplished by the proposed technique in the realm of UAV-based small-object detection, residual limitations persist. For example, the multi-branch structure, while enhancing feature representation, inevitably introduces additional computational overhead and increases the parameter count, which may limit deployment on highly resource-constrained UAV platforms. In addition, the current cross-scale attention mechanism primarily focuses on local inter-scale feature interactions, leaving room for improvement in modeling long-range cross-scale dependencies. Future work will focus on lightweight designs for multi-branch structures, more efficient cross-scale attention modeling strategies, and enhancing robustness in extremely complex scenarios, with the goal of further advancing the practicality and real-world applicability of UAV small-object detection methods.