EMWMS-YOLO: Efficient Multi-Scale Detection Framework for Small Objects in Challenging Remote Sensing Scenes

Highlights

What are the main findings?

• To address the challenges of multi-scale target detection, we propose an Efficient Multi-Scale Cross-Layer Extraction (EMSCLE) backbone framework. By effectively capturing and leveraging multi-scale features, the network enables comprehensive fusion of hierarchical information, thereby enhancing its representation capability in complex multi-scale detection scenarios.
• To address the challenges posed by complex background interference, a Multi-Scale Adaptive Feature Fusion (MSAFF) neck architecture is introduced. By performing efficient multi-scale feature fusion and incorporating an adaptive channel enhancement mechanism, the model selectively strengthens responses to informative features while suppressing redundant ones, thereby improving its capability to detect small objects in complex background scenarios.
• We design a set of core modules, including Dual-Branch Feature Extraction (DBFE), Multi-Scale Feature Perception (MSFP), Spatial Pyramid Pooling Fast with Large Separable Kernel Attention (SPPF-LSKA), Channel-Enhanced Convolution (CEC), Multi-Scale Gated Feature Fusion (MSGFF), and WaveletPool, along with the Detect-MultiSEAM detection head. These components collectively enhance the model’s ability to perceive, represent, and interact with multi-scale features, thereby enabling more effective detection of small-scale targets in complex scenes.
• We introduce the ShapeIoU loss function, which effectively addresses the insensitivity of standard IoU to the localization of elongated, irregular, or small targets, thereby enabling more accurate shape fitting in detection results.

What are the implications of the main findings?

• The DBFE module employs dual convolutional branches to capture multi-scale features while enhancing cross-channel information interaction and fine-grained feature representation. The MSFP module performs multi-scale feature fusion by leveraging the complementary strengths of MANet and FasterNet-Block, thereby improving the model’s capability in multi-scale detection tasks. The SPPF-LSKA module expands the receptive field by integrating multi-scale features and introducing a large-kernel separable attention mechanism, which enhances feature representation while reducing computational redundancy. The CEC module combines MBConv and EffectiveSE to adaptively recalibrate channel-wise responses, thereby strengthening the representation of small and low-contrast targets while alleviating performance degradation caused by background interference. The MSGFF module integrates EfficientViM and CGLU to emphasize salient features and suppress redundant information, improving the network’s effectiveness in detecting small objects under complex backgrounds. The WaveletPool module utilizes wavelet decomposition to achieve information-preserving downsampling, enhancing the representation of structural and fine-detail features while reducing spatial resolution, thus benefiting small-object detection. Finally, the Detect-MultiSEAM detection head optimizes feature representation through a multi-scale spatial enhancement mechanism, improving detection accuracy while reducing missed detections and false positives.
• The proposed EMWMS-YOLO is evaluated on two benchmark datasets, VEDAI and NWPU-VHR-10. Experimental results demonstrate that it outperforms the YOLOv11n baseline by 9.8% and 4.1% in terms of mAP50, respectively, highlighting its effectiveness in small-object detection.

Abstract

Nowadays, remote sensing images are characterized by significant scale variations, a high density of small targets, and complex background conditions, which pose substantial challenges for small-object detection. To address these issues, we propose EMWMS-YOLO, a lightweight and efficient detection framework built upon YOLOv11n. Specifically, an Efficient Multi-Scale Cross-Layer Extraction (EMSCLE) backbone is designed by integrating the Dual-Branch Feature Extraction (DBFE), Multi-Scale Feature Perception (MSFP), and Spatial Pyramid Pooling Fast with Large Separable Kernel Attention (SPPF-LSKA) modules, enabling effective multi-scale feature extraction and cross-channel interaction. Furthermore, a Multi-Scale Adaptive Feature Fusion (MSAFF) neck architecture, composed of the Channel-Enhanced Convolution (CEC) and Multi-Scale Gated Feature Fusion (MSGFF) modules, is introduced to dynamically fuse cross-scale features and enhance salient target responses while suppressing background noise. In addition, the WaveletPool module replaces conventional pooling operations to reduce information loss and feature aliasing while preserving structural details. A Detect-MultiSEAM detection head is constructed by embedding a multi-scale spatial enhancement attention mechanism, which improves feature representation under complex conditions and reduces missed detections and false positives. Finally, the ShapeIoU loss function is employed to better model geometric and morphological properties, thereby improving localization accuracy. Experimental results on the VEDAI and NWPU-VHR-10 datasets demonstrate that the proposed method achieves improvements of 9.8% and 4.1% in mAP50 over the YOLOv11n baseline, respectively, verifying its effectiveness in small-object detection.

1. Introduction

As a non-contact observation method, remote sensing technology can rely on satellites, aircraft, and unmanned aerial vehicle platforms to obtain earth surface information under large-scale and multi-temporal conditions, and is extensively employed in urban development planning [1,2,3], environmental monitoring [4,5,6], disaster assessment [7,8], agricultural management [9,10], and military reconnaissance [11,12], among others. Nowadays, deep learning-based approaches, especially the YOLO framework [13,14,15,16,17,18], have demonstrated remarkable efficacy in identifying small-scale targets, benefiting from their robust feature modeling and efficient inference mechanisms. It provides new solutions for improving detection and promoting intelligent applications.

Remote sensing target identification methods powered by deep learning can generally be classified as single-stage or two-stage models. Typical examples of the two-stage approach include R-CNN [19], Fast R-CNN [20], Faster R-CNN [21], and Mask R-CNN [22]. The core operation process is as follows: The first step is to use a region proposal network to extract potential target candidate regions; in the subsequent stage, these candidate regions are optimized through two key links of category determination and bounding box adjustment. This technical solution has shown excellent recognition performance in the target detection task. However, its obvious disadvantages lie in the high computation cost and the slow reasoning process. The one-stage detection method transforms the target recognition task into an end-to-end regression problem, and at the same time performs target classification and positioning on the entire complete image, without clearly generating candidate regions. Representative algorithms include SSD [23], RetinaNet [24], and YOLO family. In contrast to the two-stage approaches, YOLO-based models have gained extensive adoption in remote sensing small-object recognition in recent years, owing to their favorable trade-off between inference speed and detection precision. For example, YOLOv5 [25,26] introduces a decoupled detection head and an efficient feature fusion structure, YOLOv8 [27] further adopts an anchor-free strategy and lightweight modules, and the latest YOLOv10/YOLO11 [28,29] further breaks through the balance between detection efficiency and accuracy through large-kernel convolution, spatial-channel decoupling and improvement of the attention mechanism.

Although remote sensing object detection has achieved significant progress, numerous challenges still hinder its broader application and further optimization. Due to variations in imaging altitude, sensing platforms, and observation conditions, remote sensing images exhibit substantial scale variations and high intra-class diversity; in particular, small targets often occupy only a limited number of pixels, making them prone to the loss of fine-grained features during network downsampling, which consequently degrades detection accuracy. In addition, remote sensing scenes are typically characterized by complex backgrounds and various sources of interference, and the high similarity in texture and spectral features among different land-cover types can easily lead to false positives and missed detections. Meanwhile, small targets often present blurred boundaries, weak textures, and limited sample availability in real-world imagery, further increasing the difficulty of feature extraction and classification. Moreover, remote sensing imagery is frequently affected by variations in viewing angles, illumination conditions, noise interference, and geometric distortions, all of which pose significant challenges to the robustness of detection models. Furthermore, when deployed on resource-constrained platforms such as unmanned aerial vehicles (UAVs) and satellites, object detection algorithms must balance detection accuracy with real-time performance, thereby imposing stricter requirements on lightweight architecture design and computational efficiency. Therefore, achieving accurate and robust detection of multi-scale small targets in complex environments, while maintaining high efficiency, remains a critical challenge in remote sensing object detection.

Among the YOLO family, selecting an appropriate baseline model is critical for achieving a balance between detection accuracy and computational efficiency. In this work, YOLOv11n is chosen as the baseline due to its lightweight architecture, reduced parameter size, and high inference speed, which make it particularly suitable for deployment in resource-constrained remote sensing scenarios. Compared with larger variants, YOLOv11n provides a more efficient framework with lower computational cost while maintaining competitive performance. In this work, YOLOv11n is adopted as the baseline due to its lightweight architecture and fast inference speed, making it suitable for resource-constrained remote sensing scenarios. However, YOLOv11n still suffers from several limitations in remote sensing small object detection. Specifically, the conventional downsampling strategy leads to the loss of fine-grained details of small objects, the feature extraction process contains redundant spatial and channel information, the multi-scale feature fusion capability is insufficient for handling objects with large scale variations, and the model lacks effective mechanisms to capture high-frequency details and complex shape information in cluttered backgrounds.

In this work, we introduce EMWMS-YOLO, a newly designed architecture aimed at mitigating the aforementioned limitations, developed on the YOLO11n framework. Firstly, the DBFE module enhances the backbone’s capability for feature representation and fine-grained detail extraction, the MSFP module further strengthens the network’s adaptability to complex environments through its modular design, and the SPPF-LSKA module improves the feature extraction efficiency and target perception ability on the basis of reducing redundant calculations. Secondly, the CEC module effectively alleviates problems like gradient disappearance and explosion during deep network training. This module achieves more efficient feature representation while enhancing hierarchical feature association and robustness. The MSGFF module boosts the adaptive modeling of inter-feature dependencies, enabling the network to suppress redundant information under complex backgrounds or target appearance variations, thereby enhancing the recognition capability for small targets and key regions. Moreover, the WaveletPool module transforms the traditional direct compression method into a controllable downsampling method based on frequency decomposition. This reduces information loss and feature aliasing without significantly increasing parameters, enhances the representation of small targets and detailed features, and thus improves detection performance. Then, the Detect-MultiSEAM detection head effectively suppresses missed detections and false detections in small-target detection, improves detection accuracy, and reduces errors caused by occlusion or background interference. Finally, integrating ShapeIoU allows the model to maintain superior stability and accuracy when processing objects with large aspect ratio variations or irregular geometries, substantially enhancing its ability to capture geometric characteristics of targets.

In this research, the main contributions can be summarized as follows:

• An EMSCLE is proposed. This method integrates the DBFE, MSFP and SPPF-LSKA modules, which promotes the propagation of fine-grained information and enables the model to maintain a strong feature recognition ability when facing scenes of small targets or cluttered backgrounds. Moreover, by effectively sensing and utilizing multi-scale features, the network achieves the comprehensive fusion of hierarchical information, which greatly improves the representation ability in multi-scale detection tasks.
• MSAFF is created, which is composed of CEC and MSGFF components. MSAFF strengthens the model’s capability to interact with multi-scale features, enabling it to maintain efficient reasoning while keeping a balance between local details and global semantic representation. Therefore, the proposed neck network improves the detection accuracy and stability, and also significantly enhances the feature discrimination ability under complex background conditions.
• The WaveletPool module is proposed, which decomposes the frequency of features during the downsampling process through discrete wavelet transform, separates the structural information from the detail information, and selectively retains the key sub-bands in the compression stage, so as to reduce the information loss while reducing the resolution and computational cost, and enhance the integrity and discrimination of features, which is conducive to the preservation of small targets and edge details, and provides a more stable and effective feature representation for subsequent detection.
• The MultiSEAM module is embedded into the YOLO11n detection head to form the Detect-MultiSEAM detection head. Detect-MultiSEAM, combined with the Channel–Spatial Mixture Module (CSMM), effectively suppresses missed detections and false detections in small-target detection. The CSMM uses parallel convolutional kernels to capture multi-scale information, combines depthwise separable convolution and pointwise convolution to enhance feature interaction, and strengthens channel information through global average pooling and fully connected networks, thereby improving detection accuracy and reducing errors caused by occlusion or background interference.
• The ShapeIoU loss is incorporated into the framework, which applies shape-aware constraints based on the conventional IoU metric. The metric assesses both the intersection between the predicted and ground-truth bounding boxes and the consistency of object contours and geometry. In addition, it effectively compensates for the insensitivity of standard IoU to the localization of elongated, irregular, or small targets, leading to more accurate shape fitting in detection results.

2. Related Work

2.1. Multi-Scale Target Identification Methods

Multi-scale object detection methods in remote sensing imagery are developed to overcome the inconsistent detection performance caused by differences in target scale. Guo et al. [30] proposed that incorporating a cross-attention mechanism between high- and low-resolution feature maps enables the effective transfer of semantic information from high-resolution representations to low-resolution ones, thereby facilitating a more efficient fusion of multi-scale features. Teng et al. [31] introduced the MSP module, in which the acquired global contextual information is integrated with the local spatial dependencies modeled by a Clip-LSTM network. Using rich semantic representations, an adaptive anchoring mechanism is subsequently developed to mitigate the impact of target scale variations. Li et al. [32] proposed an improved feature pyramid network framework, where feature maps are progressively generated in a top-down path to achieve accurate detection of objects across multiple scales. Wang et al. [33] achieved robust features in this network through the integration of spatial mixing and efficient multi-scale attention mechanisms The trade-off between representation and computational efficiency enhances feature learning and can accurately detect objects at different scales. Yuan et al. [34] used a bilateral attentional feature mixer to integrate these features to achieve efficient fusion and enhance multi-scale target features in response to target scale changes.

The current algorithm has made significant improvements in the retrieval and augmentation of multi-scale characteristics. However, existing algorithms need to be enhanced to address the challenges posed by target scale variations, uneven feature expression, and the sensitivity of small targets to background interference.

2.2. Algorithms for Detecting Objects with Complex Backgrounds

To effectively mitigate the impact of cluttered backgrounds in remote sensing imagery. Ni et al. [35] proposed a prototyping learning network with schematic relationships, called PLNet-PR, for target detection while enhancing the sensory field, which is beneficial to identify small objects in remote sensing imagery under challenging environmental conditions. Yang et al. [36] developed an automated method for detecting aircraft targets by integrating convolutional neural networks with multi-scale cyclic frequency filtering frameworks. The designed filters effectively reduce image noise and enable the extraction of potential aircraft areas at various spatial resolutions. In Ma et al.’s study [37], the interference among distinct feature categories was mitigated by partitioning the feature map into several channels and isolating category-specific features. This process effectively suppresses background noise and enhances detection precision. Gao et al. [38] developed SBM together with MES to mitigate the disturbances introduced by complex background environments. Gan et al. [39] proposed a semi-parallel cooperative attention framework to jointly capture both global and local representations, effectively mitigating the challenges of ship feature extraction under complex background interference.

Existing background suppression approaches frequently concentrate exclusively on eliminating background noise or merely strengthening target characteristics to attain suppression outcomes. However, the issue of concurrently suppressing background interference while boosting the prominence of the target remains to be addressed.

2.3. Efficient Convolutional Network

Which further strengthens the identification of small-scale targets and detailed characteristics. Liang et al. [40] designed an edge–cloud collaborative object detection framework named Edge YOLO by reconstructing convolutional neural networks to integrate pruned feature extraction modules with compressed feature fusion components. This lightweight architecture enhances the efficiency of multi-scale prediction. Zhang et al. [41] used a multi-branch convolutional structure combined with extended convolution, which improves the ability of the network to capture and distinguish semantic information. Fan et al. [42] significantly improved the precision of target identification by introducing large-core depthwise separable convolutions to obtain global context information. Wu et al. [43] replaced the standard convolution operation with a lightweight SEConv module to decrease the network’s parameter count. This substitution significantly accelerates the detection procedure. Wu et al. [44] introduced multi-scale and multi-branch cross-scale convolutional fusion modules, combining window-based multi-head self-attention with dynamic transformation layers. This integration greatly strengthens the network’s capacity for feature learning.

Nevertheless, most existing methods primarily concentrate on extracting global features, and the limited awareness of local details leads to insufficient exploitation of fine-grained target information, thereby constraining the improvement of model performance in detecting small objects.

3. Material and Methods

3.1. Overall Framework

The Ultralytics team released the YOLO11 architecture in 2024, as shown in Figure 1. The YOLO11 series includes a total of five model versions in different scales: YOLO11n, YOLO11s, YOLO11m, YOLO11I and YOLO11x. The main differences are reflected in the depth, width and maximum number of channels of the network, which leads to differences in the computational amount and parameter scale. Its overall architecture is mainly made up of a backbone, neck and head. It enables real-time object detection with lower latency compared to other YOLO versions, and with an enhanced multi-scale feature extraction mechanism, it can effectively detect a wide range of scale targets. YOLO11’s frameless design allows for more flexibility in the model, eliminating the reliance on predefined anchor frame configurations, reducing model complexity and improving detection capabilities for diverse targets. Although YOLO11 performs well in most scenarios, it can still miss or falsely detect in complex backgrounds, multi-target occlusion, and very small targets, as well as display insufficient accuracy. To mitigate these limitations, we put forward a network structure based on the YOLO11n algorithm, namely EMWMS-YOLO. Specifically, within the backbone network, we first proposed EMSCLE, which combines dual convolution (DualConv) [45], Mixed Aggregation Network (MANet) [46] and FasterNet-Block [47]. Secondly, in the neck network part, we propose MSAFF, which combines MBConv [48], EffectiveSE [49], EfficientViMBlock [50], and CGLU [51]. Finally, ShapeIoU is employed to strengthen the awareness of target boundary dimensions and contours, thereby improving the accuracy of box position prediction. The complete framework is illustrated in Figure 2.

3.2. EMSCLE Backbone Network

Owing to the relatively small object size in remote-sensing imagery and the abundance of fine-grained spatial details, the backbone network often struggles to effectively recognize objects appearing at diverse spatial scales. To alleviate this limitation, we developed the EMSCLE backbone network, which is made up of three main parts. One of them is the DBFE module, which employs dual convolutional operations to capture representations across multiple scales simultaneously, enhancing the cross-channel communication of features and improving the model’s strength in capturing details. The MSFP module effectively aggregates multi-scale representations by integrating the strengths of Mixed Aggregation Network (MANet) and FasterNet-Block, thereby enhancing the model’s capability in handling multi-scale scenarios. The SPPF-LSKA module enhances the ability to fuse multi-scale features and information representation by embedding LSKA into the SPPF module, thereby expanding the receptive field, improving target detection performance, and reducing computational overhead while increasing operational efficiency.

3.2.1. DBFE Module

To enhance the EMWMS-YOLO backbone’s capacity for feature extraction and fine-grained information acquisition, this chapter proposes the DBFE module. DBFE optimizes the grouped transmission of information flow, can simultaneously obtain spatial clues from the feature representation and carry out cross-channel information fusion so that it can enhance the integration precision for small objects, and can also reduce information loss, maintain high accuracy in noisy and low-contrast environments, and at the same time take into account the stability. The DBFE module is as shown in Figure 3.

The DBFE block comprises a pair of dual convolutional operations (DualConv), which partition N convolutional kernels into G independent groups, each processed separately. The overall feature representation serves as the module’s input. Within each group, (M/G) input channels are convolved with both 3 × 3 and 1 × 1 filters, while the remaining (M − M/G) channels are handled using only 1 × 1 convolutions. After aggregating the outputs of 3 × 3 and 1 × 1 filters, DualConv significantly reduces the parameter count of the original backbone by applying a grouped convolution scheme and facilitates inter-layer feature communication by retaining the initial feature information from the input representation. This design captures local texture information at a shallow level while extracting global contextual features at a deeper level, achieving complementary integration of fine-grained details and advanced semantics, while also enhancing cross-channel communication capabilities. Especially in remote sensing images, small objects are challenging to identify owing to low resolution and cluttered backgrounds. The dual-branch design of DualConv effectively mitigates feature redundancy and minimizes information degradation common in single-layer convolution, thereby improving the accurate extraction of small objects and edge structures through multi-scale feature learning and representation acquisition, while ensuring robustness in environments with large noise interference.

Let the output feature map have the dimensions of D₀ × D₀ × N. In the DualConv layer composed of G groups of convolution kernels, the overall count of floating-point computations (FLOPs) related to the combined convolution kernels is as follows:

$$\mathrm{F}{\mathrm{L}}_{\mathrm{C}\mathrm{C}}={\displaystyle \frac{\left({\mathrm{D}}_0^2\times {\mathrm{K}}^2\times \mathrm{M}\times \mathrm{N}+{\mathrm{D}}_0^2\times \mathrm{M}\times \mathrm{N}\right)}{\mathrm{G}}}$$

(1)

D₀ denotes the spatial width and height of the output feature map. N denotes the total number of convolutional filters, M denotes the total number of input feature channels, and K × K defines the spatial dimensions of the convolutional kernel.

The computational cost associated with the remaining 1 × 1 pointwise convolution can be formulated as:

$${\mathrm{F}\mathrm{L}}_{\mathrm{P}\mathrm{C}}=\left(\begin{array}{c}{\mathrm{D}}_{\mathrm{o}}^2\times \mathrm{M}\times \mathrm{N}\end{array}\right)\times \left(1-{\displaystyle \frac{1}{\mathrm{G}}}\right)$$

(2)

The overall computational cost is as follows:

$$\begin{array}{rr}\hfill \mathrm{F}{\mathrm{L}}_{\mathrm{D}\mathrm{C}}& =\mathrm{F}{\mathrm{L}}_{\mathrm{C}\mathrm{C}}+\mathrm{F}{\mathrm{L}}_{\mathrm{P}\mathrm{C}}\hfill \\ \hfill & ={\mathrm{D}}_{\mathrm{o}}^2\times {\mathrm{K}}^2\times \mathrm{M}\times {\displaystyle \frac{\mathrm{N}}{\mathrm{G}}}+{\mathrm{D}}_{\mathrm{o}}^2\times \mathrm{M}\times \mathrm{N}\hfill \end{array}$$

(3)

The overall computational cost incurred by a standard convolutional layer, denoted as FLSC, can be formulated as:

$${\mathrm{F}\mathrm{L}}_{\mathrm{S}\mathrm{C}}={\mathrm{D}}_{\mathrm{o}}^2\times {\mathrm{K}}^2\times \mathrm{M}\times \mathrm{N}$$

(4)

By contrasting the computational costs between the dual convolutional layer and the standard convolutional layer, the resulting reduction ratio RDC/SC can be expressed as:

$${\mathrm{R}}_{{\displaystyle \frac{\mathrm{D}\mathrm{C}}{\mathrm{S}\mathrm{C}}}}={\displaystyle \frac{\mathrm{F}{\mathrm{L}}_{\mathrm{D}\mathrm{C}}}{\mathrm{F}{\mathrm{L}}_{\mathrm{S}\mathrm{C}}}}={\displaystyle \frac{1}{\mathrm{G}}}+{\displaystyle \frac{1}{{\mathrm{K}}^2}}$$

(5)

From the above equation, it can be seen that when selecting a larger G in the DualConv module, the DualConv module’s computing speed can be increased with the appropriate increase in the convolutional kernel K.

3.2.2. MSFP Module

Mixed Aggregation Network (MANet) combines multiple feature extraction techniques to effectively extract feature information from different levels. Its advantage is that through cross-layer feature perception and information capture, it improves the model’s capability to identify targets across different scales, which is conducive to promoting the integration of spatial representations and improving the model’s perception of target boundaries and details. Although MANet effectively improves the feature expression ability and cross-layer information interaction efficiency through the multi-scale feature extraction strategy, its structural design still has certain limitations, and there are still problems of feature redundancy and optimization instability in high-resolution input or small-target-dense scenarios. To address these limitations, we developed the MSFP module accordingly. MSFP is constituted by a 1 × 1 bypass convolution used to adjust features across channels, a depthwise separable convolution for efficient spatial representation processing, and a FasterNet block for refined hierarchical feature enhancement, which not only combines the advantages of MANet, but also optimizes the feature transfer path by introducing a partial convolution mechanism, making the information flow more efficient and coherent, thereby improving the original MANet and the problem of feature redundancy in structural design. At the same time, its anti-residual design and cross-layer feature reuse mechanism enhance the stability of information transmission and the fine-grained expression of fine-scale object characteristics, which further enhances the model’s detection precision and robustness, while preserving the merits of multi-level representation. MSFP is illustrated in Figure 4.

The FasterNet block component is constituted by a PConv layer together with two 1 × 1 convolutional layers, forming an anti-residual structure. PConv applies convolution operations to a selected portion of the input feature channels, which enhances the transmission of feature information and facilitates smoother and more coherent information flow within the network. Subsequent 1 × 1 convolutions further refine the extracted features and strengthen the information flow by introducing an intermediate activation function along with a batch normalization operation. Finally, the input features are reutilized through skip connections, ensuring the continuity of information propagation and the perception of multi-scale features.

The floating-point operation (FLOP) and memory consumption of PConv are illustrated in Equations (6) and (7), where h and w represent the feature map’s height and width, and k indicates the dimension of the convolution filter.

$$\mathrm{h}\times \mathrm{w}\times {\mathrm{k}}^2\times {\mathrm{c}}_{\mathrm{p}}^2$$

(6)

$$\mathrm{h}\times \mathrm{w}\times 2{\mathrm{c}}_{\mathrm{p}}+{\mathrm{k}}^2\times {\mathrm{c}}_{\mathrm{p}}^2\approx \mathrm{h}\times \mathrm{w}\times 2{\mathrm{c}}_{\mathrm{p}}$$

(7)

3.2.3. SPPF-LSKA Module

The SPPF module in the YOLO11 network can effectively capture the multi-scale features of the image through spatial pyramid pooling, greatly improve the feature extraction ability of the backbone, and provide rich feature support for remote sensing small-object detection. However, the redundancy in the processing process of the SPPF module will reduce the computational efficiency and running speed of the model when processing remote sensing images, so we propose the SPPF-LSKA module to solve this problem. The SPPF-LSKA module first processes the input features through 1 × 1 convolution. Subsequently, a three-layer cascaded 5 × 5 max pooling layer is used. The outputs of these pooled layers are concatenated to form a rich multi-scale representation. The aggregated tensor is then fed into the LSKA module. First, a nucleus of size K × K is decomposed into (2D − 1) × (2D − 1) deep convolution, K/d × K/d adjustable expansion rate deep expanded convolution, and 1 × 1 convolution. Subsequently, the two-dimensional deep convolutional kernel and the deep-extended convolutional kernel are further decomposed into one-dimensional horizontal convolutional kernels and vertical convolutional kernels. Finally, the decomposed convolutional kernels are connected sequentially in series.

The improved SPPF-LSKA can expand its perception range and improve its target monitoring ability by collecting feature images of different scales and extracting richer feature information, effectively reducing computational complexity and improving operational efficiency. The SPPF-LSKA is illustrated in Figure 5.

3.3. MSAFF Neck Network

To mitigate the decline in detection accuracy for small targets in complex backgrounds and large changes in target morphology, we propose the MSAFF Neck network. MSAFF consists of two core modules: the CEC module and the MSGFF module. The efficient integration and dynamic weight distribution of cross-scale features are realized through an efficient feature fusion mechanism and adaptive channel enhancement strategy. The CEC module adaptively recalibrates the weight of each channel by integrating MBConv and EffectiveSE, thereby strengthening the feature representation of small or low-contrast targets as well as reducing accuracy loss due to background interference. MSGFF can adaptively enhance the response of significant features while reducing the influence of redundant ones by combining EfficientViM and CGLU, which enhances the network’s capability to detect small targets in complex contexts.

3.3.1. CEC Module

The CEC module is a module integrating deep separable convolution and channel attention mechanism, which aims to keep feature information flowing and improve the diversity of feature representations by decoupling the feature learning of space and channel dimensions, and adaptively recalibrating channel weights within the feature map so as to boost the detection accuracy for small-scale targets when complex background and target morphology change greatly. Built upon the C3K2 framework, CEC significantly enhances the noise suppression ability by integrating MBConv deep separable convolution and the EffectiveSE channel attention mechanism. MBConv can effectively decouple the spatial and channel dimensions, mitigate noise interference and enhance the identification performance for low-contrast targets. The EffectiveSE mechanism strengthens the feature response related to the target key point through adaptive channel weight allocation. The CEC is illustrated in Figure 6.

MBConv combines 1 × 1 pointwise convolution with depthwise separable convolution to efficiently capture the feature representations. Initially, it expanded the input channel with 1 × 1 convolution to expand the feature dimension. Subsequently, the spatial information is extracted by deep separable convolution. Squeeze-and-Excitation is used to strengthen the expression of key features via the channel attention mechanism, and another 1 × 1 convolution is applied. Convolution performs feature integration and compression. Finally, the residual pathway is introduced to promote gradient propagation and improve training efficiency. Such a design greatly improves the stability and wide adaptability of the network, making it both efficient and scalable, and able to adapt to application scenarios such as high-precision image recognition and target detection.

The EffectiveSE block is composed of extrusion and excitation. In the extrusion phase, global average pooling is performed to extract the overall contextual information of all channels. In the Excitation phase, a multi-layer perceptron is employed to determine channel-specific weights. The scalar output from the Squeeze operation is then passed through a fully connected network, which calculates the channel weights. The obtained weights are subsequently multiplied with the original feature map on an element-by-element basis, producing a reweighted feature map that strengthens the representation of important features. The mathematical formulation of EffectiveSE is given in Equations (8)–(10).

First, for the input feature graph ${\mathrm{X}}_{\mathrm{i}\mathrm{j}\mathrm{c}}$, the weights of each channel are pooled globally:

$${\mathrm{W}}_{\mathrm{c}}={\displaystyle \frac{1}{\mathrm{H}\times \mathrm{W}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{H}}}{\displaystyle \sum _{\mathrm{j}=1}^{\mathrm{W}}}{\mathrm{X}}_{\mathrm{i}\mathrm{j}\mathrm{c}}$$

(8)

The weight w_c is then transformed into an updated activation value f′(w_c; θ) (where θ is a learnable parameter) and sigmoid activation of the activation value:

$${\mathrm{S}}_{\mathrm{c}}=\mathsf{\sigma }\left({\mathrm{f}}^{\prime }\left({\mathrm{w}}_{\mathrm{c}};\mathsf{\theta }\right)\right)$$

(9)

Finally, the weighted S_c of each channel is multiplied by the original feature representation X_ijc to generate a reweighted feature map Y_ijc.

$${\mathrm{Y}}_{\mathrm{i}\mathrm{j}\mathrm{c}}={\mathrm{S}}_{\mathrm{c}}{\mathrm{X}}_{\mathrm{i}\mathrm{j}\mathrm{c}}$$

(10)

3.3.2. MSGFF Module

In the YOLO11n model, the C3K2 block uses complex convolutional operations in the feature extraction process, which greatly increases the computational complexity. However, the C3K2 module has certain limitations in feature fusion at different scales, particularly under complex background conditions or large changes in target morphology. It cannot fully and effectively fuse multi-scale information, affecting identification precision. To overcome these problems, EfficientViMBlock is proposed to replace the residual blocks in the original C3K2 structure. EfficientViMBlock enhances global feature capture by introducing an HSM-SSD layer and enables more efficient contextual modeling in multi-stage feature interactions. This structure optimizes the feature transfer path while alleviating the bottleneck of feature expression at the SSD layer, ensuring the coherence and stability of information flow. Meanwhile, the multi-stage fusion (MSF) mechanism effectively improves the synergistic expression between shallow and deep features, so that the network can show stronger feature integration and representation capabilities in complex visual scenes. However, when processing large-size inputs, the model may still be limited by memory bandwidth, and the multi-stage fusion structure makes the overall network design more complex, increasing the stability challenge of the training process. To overcome these challenges, an enhanced MSGFF module was developed. MSGFF replaces the traditional feedforward network with CGLU, which optimizes the feature transmission path through cross-level partial connection, achieves more comprehensive multi-scale feature aggregation and global context enhancement, and achieves a more efficient overall architecture. CGLU combines gating mechanisms and deep convolution to enhance feature selectivity and generalization performance while maintaining expression capabilities compared to traditional FFN, which improves the stability and adaptability of the module across diverse environments. The MSGFF module is shown in Figure 7.

The CGLU (Gated Linear Unit) module enhances local spatial awareness by introducing 3 × 3 deep convolution before traditional GLU gated branch activation functions while maintaining low computational overhead. The CGLU module consists of two branches: one captures local information through deep convolution and gating activation functions, and the other generates a gated signal and multiplies it with the output of the first branch for adaptive feature weighting. Finally, the gated signal is combined with the initial input through the residual path to produce the ultimate result. CGLU uses the local feature information extracted by deep convolution to adjust the gating signal, enhance the ability to focus on salient features, attenuate redundant features, and enhance the efficiency of the module.

3.4. WaveletPool Module

Traditional downsampling often discards fine-grained texture details and edge information, weakening the representation of small object features and reducing detection accuracy. To improve this, we propose the WaveletPool module to replace the traditional downsampling convolution in YOLO11n. As shown in Figure 8, WaveletPool achieves multi-scale decomposition and reconstruction of features through Discrete Wavelet Transform (DWT) and Inverse Discrete Wavelet Transform (IDWT). During forward propagation, the input image (or feature map) first undergoes a first-level Discrete Wavelet Transform (DWT), decomposing it into four sub-bands: a low-frequency approximation sub-band LL1 and the high-frequency sub-bands LH1, HL1, and HH1, with detailed information in different directions. The LL1 sub-band mainly retains the overall structural information of the image, while the other sub-bands characterize the details in the horizontal, vertical, and diagonal directions, respectively. Subsequently, a second-level Discrete Wavelet Transform (DWT) is performed on the low-frequency sub-band LL1 to generate four new sub-bands: LL2, LH2, HL2, and HH2, to extract coarser-scale global structural features, thus forming a hierarchical multi-scale representation. In the feature compression stage, some high-frequency sub-bands are discarded or merged according to network design requirements to retain as much effective information as possible while reducing feature resolution and computational complexity. Finally, the retained sub-bands are reconstructed using inverse discrete wavelet transform (IDWT) to achieve feature downsampling and information integration. Compared with traditional pooling or stride convolution, wavelet pooling has reversibility and clear frequency decomposition characteristics throughout the downsampling process, which can effectively alleviate information loss and feature aliasing problems. While maintaining global structural information, it strengthens edge and detail representation, providing richer and more discriminative feature representations for subsequent detection networks.

3.5. Detect-MultiSEAM Detection Head

In small-target detection, the target is often obscured by other objects, which may lead to the disappearance of features, resulting in inaccurate positioning and affecting detection accuracy. Therefore, we innovatively propose to embed the MultiSEAM module into the YOLO11n detection head to form the Detect-MultiSEAM detection head, which significantly inhibits the missed detection and false detection in the detection of small targets, and improves the detection accuracy of the model.

The design of the MultiSEAM module is shown in Figure 9. As shown, the principal component of it is the Channel–Spatial Mixture Module (CSMM), which processes input features through parallel convolutional kernels at different scales to capture multi-scale information more comprehensively and achieve more accurate defect detection and localization. In the CSMM, depth-separable convolutions are combined with residual links to enable cutting calculations while maintaining feature authenticity. Since processing channel information alone may overlook the correlation between channels, point-by-point convolution (1 × 1 convolution) is used to integrate channel features and enhance feature interaction. The features output by the CSMM first undergo global average pooling to compress spatial resolution, eliminate unnecessary redundancy, and strengthen feature consistency. Next, a two-stage fully connected network consolidates channel-wise representations to reinforce interactions across different feature maps. Finally, channel expansion combines information from multiple scales, effectively mitigating detection inaccuracies arising from occlusions or complicated background effects.

3.6. Loss Function ShapeIoU

CIoU extends traditional IoU by considering the overlap between predictive bounding boxes and ground reality bounding boxes. Other factors are also taken into account, such as the distance between their centers, aspect ratio, and rotation angle. This enables the network to improve target positioning accuracy during bounding box prediction, enabling more accurate target positioning, especially in small-target detection. Although CIoU has improved the positioning accuracy of bounding box regression, there are some limitations. First, when the target has obvious overlap or severe shape change, CIoU may not be able to produce the optimal regression performance. Moreover, CIoU ignores the geometric characteristics of the target and does not fully consider the impact of the scale and shape of the bounding box on the regression result. In order to overcome these deficiencies, ShapeIoU is introduced, which emphasizes the scale and geometric configuration of the bounding box, thereby improving the precision of bounding box regression. The construction of ShapeIoU is as shown in Figure 10.

The mathematical formula of ShapeIoU:

$$ww={\displaystyle \frac{2\times {\left(w^{\mathrm{g}\mathrm{t}}\right)}^{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}}}{{\left(w^{\mathrm{g}\mathrm{t}}\right)}^{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}}+{\left(h^{\mathrm{g}\mathrm{t}}\right)}^{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}}}}$$

(11)

$$hh={\displaystyle \frac{2\times {\left(h^{\mathrm{g}\mathrm{t}}\right)}^{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}}}{{\left(w^{\mathrm{g}\mathrm{t}}\right)}^{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}}+{\left(h^{\mathrm{g}\mathrm{t}}\right)}^{\mathrm{s}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e}}}}$$

(12)

$${distance}^{shape}=hh\times {\displaystyle \frac{{\left({\mathrm{x}}_{\mathrm{c}}-{\mathrm{x}}_{\mathrm{c}}^{\mathrm{g}\mathrm{t}}\right)}^2}{{\mathrm{c}}^2}}+ww\times {\displaystyle \frac{{\left({\mathrm{y}}_{\mathrm{c}}-{\mathrm{y}}_{\mathrm{c}}^{\mathrm{g}\mathrm{t}}\right)}^2}{{\mathrm{c}}^2}}$$

(13)

$${\Omega }^{shape}=\sum _{\mathrm{t}={\mathsf{\omega }}_{\mathrm{a}},{\mathsf{\omega }}_{\mathrm{b}}}{\left(1-{\mathrm{e}}^{-{\mathsf{\omega }}_{\mathrm{t}}}\right)}^{\mathsf{\theta }},\mathsf{\theta }=4$$

(14)

$$w_a=hh\times {\displaystyle \frac{\left|w-w^{gt}\right|}{max\left(w,w^{gt}\right)}}$$

(15)

$$w_b=ww\times {\displaystyle \frac{\left|h-h^{gt}\right|}{max\left(h,h^{gt}\right)}}$$

(16)

Scale represents the scale factor related to the relative proportion of the target in the dataset, while ww and hh represent the weight parameters along the horizontal and vertical axes, respectively. Its size correlates with the geometry of the ground live (GT) box. The bounding box regression loss based on the proposed ShapeIoU formulation is defined as follows:

$$L_{ShapeIoU}=1-IoU+{distance}^{shape}+0.5\times {\Omega }^{shape}$$

(17)

4. Experimental Part

4.1. Experimental Dataset

The VEDAI dataset [52] is an aerial imagery dataset with two available resolutions: 1024 × 1024 and 512 × 512. The dataset contains 1246 images labeled with 9 target categories: trucks, cars, pickup trucks, campers, vans, airplanes, tractors, boats, and other examples. The proportion used for data partitioning of the training, validation, and test sets is 7:1:2.

The NWPU-VHR-10 dataset [53] contains 800 high-resolution images, whose spatial resolution varies between 0.5 m and 2 m, mainly from Google Earth and aerial photography. This dataset is divided into two parts: the positive sample part has 650 images containing the target, and the negative sample part has 150 images showing only the background. This data set covers 10 target categories: ships, aircraft, storage tanks, tennis courts, blue courts, baseball fields, tracks and fields, bridges, ports, and vehicles. The proportion used for data partitioning of the training, validation, and test sets is 7:1:2.

4.2. Experimental Details

The experimental operation environment of this study was built on a remote computing node equipped with the Ubuntu 22.04 operating system. The server hardware configuration includes a twelve-core AMD EPYC 7352 central processing unit and an NVIDIA GeForce RTX 4090 graphics acceleration card with 24 GB GDDR6 video memory, 124 GB system memory and 100 GB local storage capacity. In the model development and training stage, researchers adopted the latest stable version, 2.3.0, of the PyTorch deep learning framework and combined it with the CUDA 12.1 parallel computing platform to improve the GPU operation efficiency. In addition, the 3.10.15 version of the Python programming language was used as the basic development environment, and Ultralytics YOLOv8.3.9 was selected as the specific implementation framework of the object detection model. For detailed information of the various specific parameter settings in the model training process, please refer to

Table 1. Implementation Details.

Parameter Type	Specified Parameter Values
Momentum	0.937
Weight decay	0.0005
Input image size	640 × 640
Total epochs	300
Batch size	32
Optimizer	SGD
Initial learning rate	0.01
Cache	False
Close_mosaic	0
Workers	8

of this article.

4.3. Assessment Metrics

The performance of the model is assessed through the following evaluation indicators: precision (P), recall (R), mean average precision (mAP), and million parameters (M).

$$\mathrm{P}=-{\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(18)

$$\mathrm{R}=-{\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(19)

$$\mathrm{A}\mathrm{P}={\int }_0^1\mathrm{P}\left(\mathrm{R}\right)\mathrm{d}\mathrm{R}$$

(20)

$$\mathrm{m}\mathrm{A}\mathrm{P}={\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle \sum _{\mathrm{l}=1}^{\mathrm{N}}}\mathrm{A} \mathrm{P}\left(\mathrm{i}\right)$$

(21)

In this case, the terms true positive (TP), false positive (FP) and false negative (FN) are defined. mAP is a common metric used to evaluate the performance of object detection models. We mainly used the mAP50 for performance evaluation.

4.4. Comparative Experiments

4.4.1. Evaluation of Different Backbone Network Designs

To verify the feasibility of EMSCLE, we compared LSKNet [54], Efficientvit [55], fasternet [47], Unireplknet [56] and HGNetV2 [57]. The proposed EMSCL module was incorporated into the baseline model, with results summarized in

Table 2. Backbone network comparison.

Method	P%	R%	mAP50%	Map50–90%	GFLOP	Param(M)
Baseline	59.9	55.9	58.7	33.8	6.3	2.6
+LSKNet	51.4	48.2	46.7	25.6	18.2	5.6
+Efficientvit	53.2	49.3	49.4	28.9	7.9	3.7
+fasternet	58.3	46.8	51.1	28.6	9.2	3.9
+Unireplknet	50.0	54.3	52.6	30.1	14.1	5.8
+HGNetV2	57.5	52.2	55.7	32.7	5.7	2.1
Ours	71.4	56.7	64.3	38.2	7.4	2.9

. The precision, recall, and mAP50 of EMSCL surpass those of existing mainstream models, showing improvements of 11.5%, 0.8%, and 5.6%, respectively, relative to the baseline. Overall, EMSCL demonstrates clear advantages in multi-scale detection tasks.

4.4.2. Evaluation of Different Neck Network Designs

To verify the feasibility of MSAFF, we compared EfficientRepBiPAN [17], BIFPN [58], Gold-YOLO [59], GFPN [60] and ASF [61], as well as our proposed MSAFF, to the baseline model, and the corresponding results are provided in

Table 3. Neck network comparison.

Method	P%	R%	mAP50%	Map50–90%	GFLOP	Param(M)
Baseline	59.9	55.9	58.7	33.8	6.3	2.6
+EfficientRepBiPAN	57.6	55.7	57.3	33.4	8.1	3.4
+BIFPN	54.1	62.7	59.4	35.2	6.3	1.9
+Gold-YOLO	59.1	59.6	59.8	35.3	9.2	5.9
+GFPN	60.4	60.2	61.9	35.6	8.2	3.7
+ASF	63.9	61.2	62	34.9	7.1	2.7
Ours	62.9	62.5	64.5	38.2	6.5	2.6

. MSAFF surpasses the others across multiple metrics, with precision and recall improved by 3% and 6.6%, respectively, relative to the baseline. The mAP50 was enhanced by 5.8% compared to the baseline, substantially outperforming the other models. Overall, MSAFF markedly improves the detection ability of small targets in complex backgrounds.

4.4.3. Evaluation of Different Detection Head Designs

To verify the feasibility of the Detect-MultiSEAM detection head, we compared Dy head [62], Efficient head [63], LSCSBD head [64], LSCD head [65], and our proposed Detect-MultiSEAM detection head, and the results are shown in

Table 4. Detection head comparison.

Method	P%	R%	mAP50%	Map50–90%	GFLOP
Baseline	59.9	55.9	58.7	33.8	6.3
+Dy head	54.2	53.3	54.1	32.2	7.1
+Efficient head	54.7	53.1	55.9	33.5	5.1
+LSCSBD head	54.9	57.4	57.1	31.3	6.2
+LSCD head	58.0	60.4	60.7	34.6	7.7
+Detect-MultiSEAM	61.3	64.4	62.6	37.3	6.0

. The Detect-MultiSEAM detection head outperforms other detection heads in terms of precision, recall, and Map50–90, especially in mAP50, where it improves by 3.9% compared to the baseline, while reducing GFLOP cost by 4.8%. These results demonstrate that the Detect-MultiSEAM detection head can effectively reduce false negatives and improve the accuracy of small target detection.

4.4.4. Performance Comparison Among Different Loss Function Designs

To verify the feasibility of ShapeIoU, we compared DIoU [66], EIoU [67], GIoU [68], SIoU [69], Inner-ShapeIoU [70], Focaler-ShapeIoU [71] and ShapeIoU [72] with the base model, and the results are presented in

Table 5. Loss function comparison.

Method	P%	R%	mAP50%	Map50–90%	GFLOP	Param(M)
CIoU	59.9	55.9	58.7	33.8	6.3	2.6
DIoU	66.7	58	63.2	36.9	6.3	2.6
EIoU	61.7	59.3	61.9	35.8	6.3	2.6
GIoU	62.4	57.4	60.9	35.9	6.3	2.6
SIoU	59.0	62.2	62.0	37.1	6.3	2.6
Inner-ShapeIoU	58.8	56.8	58.8	36.6	6.3	2.6
Focaler-ShapeIoU	66.8	60.2	64.0	37.5	6.3	2.6
Ours	64.9	61.0	65.1	38.5	6.3	2.6

. ShapeIoU achieves significant gains in mean average precision (mAP50), thereby substantially improving the detection performance of EMWMS-YOLO. Furthermore, compared to IoU loss functions for other objects with different shapes, our proposed ShapeIoU has a significant advantage in mAP50. In conclusion, the proposed ShapeIoU demonstrates superior performance over other mainstream loss functions in terms of boundary regression accuracy.

4.4.5. Comparative Analysis of Various Models

To evaluate EMWMS-YOLO, we compared it with two-stage detectors (such as Faster R-CNN) and single-stage detectors (such as the YOLO series) on the VEDAI dataset. The results are as shown in

Table 6. Comparison of different models.

Method	P%	R%	mAP50%	GFLOP
Faster R-CNN [21]	69.1	57.1	62.5	186.4
Mask R-CNN [22]	71.2	59.7	63.1	240.6
Cascade-RCNN [73]	-	-	61.4	91.1
CenterNet [74]	66.7	57.4	58.7	137.4
RetinaNet [24]	67.9	55.1	61.4	299.5
YOLOv3 [75]	72.4	46.6	56.7	282.3
YOLOv6n [17]	62.1	56.4	58.2	11.8
YOLOv8n [76]	71.9	55.0	58.9	8.1
YOLOX [77]	63.1	56.7	58.4	48.1
Hyper-yolo [46]	61.4	57.9	61.2	9.2
RT-DETR [57]	73.0	47.4	61.5	222.5
HPS-DETR [78]	67.4	60.3	59.4	-
YOLO11n	59.9	55.9	58.7	6.3
OURS	66.0	64.8	68.5	6.7

. Specifically, EMWMS-YOLO’s recall and mAP50 are significantly better than those of existing mainstream methods, highlighting its superior ability to detect small targets in complex backgrounds. Although EMWMS-YOLO’s precision is slightly lower than that of Faster R-CNN and Mask R-CNN, this issue is minor given the model’s lower calculation complexity and fewer parameters. It also stays within acceptable limits, indicating that our algorithm has the advantage of more efficient performance. Compared with the YOLO series algorithms, EMWMS-YOLO maintains a high detection accuracy, achieving leading performance among algorithms of the same type. The overall gain reaches 9.8 percentage points in comparison with the baseline. The findings suggest that EMWMS-YOLO delivers excellent results for small-object detection with superior accuracy.

4.4.6. Ablation Analysis

To evaluate the EMWMS-YOLO model, an ablation study was performed on VEDAI dataset using YOLO11n as baseline for various enhanced configurations. The results are summarized in

Table 7. Ablation research.

Method	P%	R%	mAP50%	Map50–90%	GFLOP
Baseline	59.9	55.9	58.7	33.8	6.3
Baseline + DBFE	68.5	55.7	63.4	38.5	7.2
Baseline + MSFP	60.5	66.8	63.9	38.1	6.6
Baseline + SPPF-LSKA	63.0	58.7	63.3	36.4	6.5
Baseline + CEC	56.1	62.6	61.0	34.3	6.4
Baseline + MSGFF	59.4	59.1	62.2	35.9	6.4
Baseline + WaveletPool	59.1	60.7	58.9	33.9	5.4
Baseline + Detect-MultiSEAM	61.3	64.4	62.6	37.3	6.0
Baseline + EMSCLE	71.4	56.7	64.3	38.2	7.4
Baseline + MSAFF	62.9	62.5	64.5	38.2	6.5
Baseline + EMSCLE + MSAFF	61.2	67.3	66.1	39.3	7.5
Baseline + EMSCLE + MSAFF + WaveletPool	63.6	63.5	66.6	39.4	6.9
Baseline + EMSCLE + MSAFF + WaveletPool + Detect-MultiSEAM	67.5	61.2	67.3	39.5	6.7
Baseline + EMSCLE + MSAFF + WaveletPool + Detect-MultiSEAM + ShapeIoU	66.0	64.8	68.5	40.4	6.7

. From Table 7, it can be seen that each module we propose contributes to the improvement of the model, reflecting the necessity of the existence of each module. EMWMS-YOLO improves precision by 6.1%, from 59.9% to 66.0%; recall improves by 8.9%, from 55.9% to 64.8%; and mAP50 improves from 58.7% relative to the YOLO11n baseline to 68.5%, an increase of 9.8%. Although a slight increase is observed in computational overhead, the results indicate that all performance metrics of EMWMS-YOLO surpass the baseline, with a particularly notable enhancement in detection accuracy. Moreover, the network demonstrates greater advantages compared to models applying only a single or partial improvement strategy. These findings confirm that EMWMS-YOLO effectively enhances the detection performance.

4.4.7. Visualization Results

To intuitively compare the detection performance of YOLO11n and EMWMS-YOLO, we conducted visualizations on the VEDAI and NWPU-VHR-10 datasets. The results are in Figure 11 and Figure 12, Respectively. EMWMS-YOLO significantly reduces both missed detections and false positives, demonstrating a significant improvement in recognition accuracy compared to the YOLO11n algorithm.

4.4.8. Generalized Verification

To evaluate the generalization capability of EMWMS-YOLO, experiments were conducted on the NWPU-VHR-10 dataset. The dataset has the same hyperparameters and training settings. The results are shown in

Table 8. Generalization Validation.

Method	P%	R%	mAP50%	GFLOP
Faster R-CNN [21]	88.2	84.5	87.4	275.4
Mask R-CNN 22]	88.6	85.6	87.6	329.7
CenterNet [74]	87.9	85.2	88.2	137.4
RetinaNet [24]	87.5	84.2	86.9	299.5
YOLOv3 [75]	88.0	81.3	87.9	282.3
YOLOv6n [17]	88.5	81.8	90.8	11.8
YOLOX [77]	89.1	84.5	89.2	48.1
Hyper-yolo [46]	92.1	89.4	93.1	9.2
RT-DETR [57]	87.1	87.5	93.4	222.5
HPS-DETR [78]	92.7	87.8	91.5	-
YOLO11n	91.0	83.2	90.7	6.3
OURS	91.3	91.7	94.8	6.7

. The table illustrates that EMWMS-YOLO outperforms Faster R-CNN, RetinaNet, YOLOv3, YOLOv6n and other comparison algorithms, and parameters such as precision, recall, and, in particular, mAP50 have significant advantages over all the comparison algorithms, with mAP50 improving by 4.1 percentage points compared with the baseline. In conclusion, EMWMS-YOLO has shown excellent generalization capabilities.

5. Discussion

The EMWMS-YOLO algorithm proposed in this paper performs well in detecting small-scale and low-contrast targets in complex remote sensing images. By integrating the EMSCLE backbone network, including the DBFE, MSFP and SPPF-LSKA modules, the network effectively enhances multi-scale feature extraction and cross-channel information interaction capabilities. Additionally, the introduction of the MSAFF neck structure, consisting of the CEC and MSGFF modules, further enhances the model’s robustness in complex environments. Moreover, the WaveletPool module effectively reduces information loss and retains multi-scale features through frequency decomposition and controllable downsampling based on wavelet transform, improving the detection ability of small targets and detailed features. Then, the Detect-MultiSEAM detection head strengthens feature expression and occlusion robustness through the multi-scale spatial enhancement attention mechanism, effectively reducing missed detection and false detection, and improving the detection accuracy of small objects. The experimental results show that the model significantly improves the detection accuracy, which verifies the effectiveness of EMWMS-YOLO.

Despite this, EMWMS-YOLO faces challenges such as increased computational overhead and limited scope for dataset validation. The performance of the model may also be affected by dataset bias and annotation quality. Future work will focus on optimizing computational efficiency through model pruning and quantification, exploring transformer-based attention mechanisms and self-supervised learning to improve contextual understanding, and extending the framework to multimodal data such as SAR or LiDAR to enhance its robustness and generalization capabilities. Overall, EMWMS-YOLO provides a solid foundation for efficient small-target remote sensing detection, with strong potential for further optimization and real-world deployment.

6. Conclusions

In this study, we propose a novel small-object detection algorithm for remote sensing imagery, termed EMWMS-YOLO, aimed at improving detection accuracy for small targets. Specifically, EMWMS-YOLO incorporates an Efficient Multi-Scale Cross-Layer Extraction (EMSCLE) backbone, which consists of the Dual-Branch Feature Extraction (DBFE), Multi-Scale Feature Perception (MSFP), and Spatial Pyramid Pooling Fast with Large Separable Kernel Attention (SPPF-LSKA) modules, enabling effective multi-scale feature extraction. The DBFE module enhances small-target recognition through dual convolutional branches while mitigating information loss, whereas the MSFP module integrates the advantages of MANet and FasterNet-Block to strengthen spatial feature interaction and improve fine-detail perception. The SPPF-LSKA module further expands the receptive field via multi-scale feature aggregation and a large-kernel separable attention mechanism, reducing computational redundancy while enhancing feature representation. In addition, a Multi-Scale Adaptive Feature Fusion (MSAFF) neck is introduced, consisting of the Channel-Enhanced Convolution (CEC) and Multi-Scale Gated Feature Fusion (MSGFF) modules, to improve feature fusion under complex backgrounds and varying target morphologies. The CEC module, built upon the C3K2 architecture, combines MBConv depthwise separable convolution with the EffectiveSE channel attention mechanism to enhance noise suppression while improving model stability and adaptability, whereas the MSGFF module integrates EfficientViMBlock with CGLU to further strengthen feature interaction. Furthermore, the WaveletPool module achieves information-preserving downsampling via wavelet decomposition, enhancing structural and fine-detail representation while reducing feature resolution, thereby improving small-object detection performance. Additionally, the Detect-MultiSEAM detection head enhances small-target recognition and reduces missed detections and false positives through multi-scale feature enhancement and spatial attention mechanisms. Finally, the ShapeIoU loss function is introduced to emphasize the geometric and morphological properties of bounding boxes, thereby improving regression accuracy.

Experimental results on the NWPU-VHR-10 and VEDAI datasets demonstrate that EMWMS-YOLO effectively handles large variations in target scale, dense object distributions, and complex background conditions. The proposed method significantly reduces false positives and missed detections, leading to more accurate object detection in remote sensing imagery. Moreover, EMWMS-YOLO achieves superior accuracy while maintaining higher computational efficiency compared to existing methods.