YOLO-SSFA: A Lightweight Real-Time Infrared Detection Method for Small Targets

Abstract

Infrared small target detection is crucial for military surveillance and autonomous driving. However, complex scenes and weak signal characteristics make the identification of such targets particularly difficult. This study proposes YOLO-SSFA, an enhanced You Only Look Once version 11 (YOLOv11) model with three modules: Scale-Sequence Feature Fusion (SSFF), LiteShiftHead detection head, and Noise Suppression Network (NSN). SSFF improves multi-scale feature representation through adaptive fusion; LiteShiftHead boosts efficiency via sparse convolution and dynamic integration; and NSN enhances localization accuracy by focusing on key regions. Experiments on the HIT-UAV and FLIR datasets show mAP50 scores of 94.9% and 85%, respectively. These findings showcase YOLO-SSFA’s strong potential for real-time deployment in challenging infrared environments.

1. Introduction

Detecting small targets in infrared imagery is crucial for a range of applications, including remote sensing monitoring, intelligent security systems, and autonomous UAV navigation [1,2]. However, factors such as low target-to-background contrast, complex backgrounds, and sensor noise significantly hinder detection performance [3,4]. Moreover, infrared images are further affected by sensor noise, false targets, and environmental interference [5]. Therefore, improving the detection accuracy of infrared small targets in complex backgrounds remains a critical challenge. To tackle this issue, researchers have proposed a variety of enhancements in feature extraction, feature fusion, and object detection head design [6,7,8,9]. Nevertheless, current approaches still exhibit certain limitations.

Firstly, in multi-scale feature fusion, traditional methods such as Feature Pyramid Network (FPN) and Path Aggregation Network (PAN) enhance multi-scale features through top-down and bottom-up information flow. However, these approaches rely on simple feature concatenation or weighted fusion, making it challenging to accurately extract key information from small targets. To address this issue, HRNet [10] employs a parallel high-resolution feature extraction mechanism, improving the detection of small targets. The Dynamic Feature Pyramid (DFP) proposed by Zhang et al. [11] adopts a dynamic attention mechanism, enabling the model to automatically optimize feature representations for different target categories and sizes. Nevertheless, these methods still struggle to fully capture contextual information across feature layers, limiting further improvements in detection performance.

Secondly, in suppressing background noise, recent studies have primarily employed attention mechanisms to enhance the representation of key target regions. For instance, the Convolutional Block Attention Module (CBAM) [12] integrates channel and spatial attention to improve the network’s focus on essential features. The Simple Attention Module (SimAM) [13] enhances target representation by explicitly computing the response distribution of neurons. Despite these developments, current infrared small target detection approaches continue to suffer from high computational overhead and limited cross-domain generalizability. A key challenge is effectively suppressing background interference while enhancing small target feature representation.

Finally, regarding detection head optimization, YOLO series detectors (e.g., YOLOv1 [14], YOLOv4 [15], and YOLOv7 [16]) use various architectures for different scenarios. For example, the PANet detection head in YOLOv4 enhances target localization capabilities, while YOLOv5 further improves detection accuracy by refining activation functions and regularization strategies. In recent years, research on lightweight detection heads has gained significant attention. For instance, NanoDet [17] employs a more compact network architecture to accommodate edge computing devices. However, these methods still struggle to balance computational efficiency and detection precision in infrared small target tasks.

To overcome the aforementioned limitations, this paper proposes an infrared small target detection method based on Scale-Sequence Feature Fusion (SSFF) and the LiteShiftHead detection head. Specifically, the key contributions of this study can be outlined as follows:

• We designed a Scale-Sequence Feature Fusion (SSFF) neck to improve multi-scale feature representation for accurate small target detection.
• We develop LiteShiftHead, a lightweight detection head that significantly reduces computational cost, enabling deployment on resource-constrained platforms.
• We designed a Noise Suppression Network (NSN) that enhances attention to salient features by effectively mitigating interference from irrelevant background regions.

2. Related Work

In traditional deep learning-based object detection models, Feature Pyramid Networks (FPN) and Path Aggregation Networks (PANs) are two typical multi-scale feature fusion methods. FPN utilizes a simple additive fusion approach, which results in insufficient feature representation for small target areas. PAN [18], proposed by Liu et al., further enhances feature flow based on FPN. However, the feature fusion method in PAN remains relatively coarse. To improve accuracy in complex scenarios, researchers have developed novel feature fusion techniques [19,20,21,22,23]. BiFPN [24] enhances feature fusion efficiency by introducing adaptive weighting to optimize information flow. Tang et al. [25] introduced Light-YOLO, which enhances small ship detection through multi-scale feature fusion, but involves deeper backbone structures. However, these methods still fall short in complex scenarios. To address this, this paper proposes a feature fusion neck structure that employs Scale-Sequence Feature Fusion (SSFF) methods. This structure establishes stronger connections between multi-level features, thereby better capturing detailed information of the targets.

In infrared target detection, background noise often interferes with small target recognition. Traditional methods use techniques like background modeling or filtering to suppress noise [26,27,28]. The Top-Hat transform proposed by Bai et al. [29] enhances target features through morphological operations, but it is prone to noise interference. The mean smoothing filter proposed by Gao et al. [30] reduces high-frequency noise but may blur target edges [31]. MST-Net, proposed by Li et al. [32], utilizes a multi-scale Transformer to extract global context information, effectively mitigating background interference. Sun et al. [33] applied non-convex weighted tensor rank minimization for small target detection, achieving strong accuracy at the cost of higher computational load. Xiong et al. [34] presented an adaptive dynamic fusion network to address complex backgrounds, though lacking deployment considerations. Li et al.’s Adaptive Background Modeling (ABM) constructs a feature distribution model via adversarial learning but struggles with rapidly changing backgrounds [35]. To address this, this paper proposes a noise suppression network that combines channel and spatial attention to effectively enhance feature representation in the target area while suppressing background interference. It significantly improves the detection performance of infrared small targets without notably increasing computational cost.

The detection head is a crucial component of the object detection framework, responsible for classifying the output of the feature extraction network and performing bounding box regression. Early detection heads often used dense connections, such as the fully connected detection heads in Faster R-CNN [36] proposed by Ren et al. and YOLOv3 [14] by Redmon et al., but these methods are computationally expensive and unsuitable for real-time detection tasks. To improve efficiency, lightweight detection heads have been proposed [37]. For instance, Yi et al.’s Efficient Detection Head (EDH) uses depth-wise separable convolutions to minimize parameter count and computation, though at the cost of weaker feature representation and limited adaptability in complex scenarios [38]. To address this, this paper designs a lightweight detection head, LiteShiftHead, which significantly reduces the computational complexity of the model while maintaining high detection performance, making it suitable for resource-constrained devices.

Transformer-based detectors have emerged as powerful tools in infrared target detection. DETR [39] and Swin Transformer-based models [40] offer global context modeling and strong feature representation, which are beneficial for small target recognition in complex backgrounds. However, their computational cost and slower inference speed limit their deployment in real-time UAV applications. Our proposed YOLO-SSFA is designed to address this trade-off by offering competitive detection performance with significantly higher FPS and fewer parameters. In addition to transformer-based architectures, recent anchor-free detectors such as YOLOX [41] have demonstrated improved performance and training stability in object detection tasks, including infrared target detection. YOLOX eliminates the need for anchor box design by predicting center points and object dimensions directly, which can simplify training and enhance generalization. However, its computational load remains significant compared to lightweight alternatives, limiting deployment on real-time UAV platforms. In contrast, YOLO-SSFA is optimized for low-latency environments while maintaining competitive accuracy.

YOLOv11 [42] is a recent evolution in the YOLO series, designed to further improve inference speed and deployment efficiency on edge devices. Compared with earlier versions like YOLOv5 and YOLOv8, YOLOv11 emphasizes lightweight architecture while preserving detection accuracy, making it a promising baseline for real-time tasks. However, YOLOv11 still suffers from limitations in handling small infrared targets and complex backgrounds, as it lacks dedicated modules for fine-grained feature fusion and attention-based noise suppression. In this work, we build upon YOLOv11n by integrating a lightweight multi-scale fusion neck (SSFF), a denoising attention module (NSN), and an efficient detection head (LiteShiftHead), aiming to enhance both accuracy and real-time performance in infrared small target detection scenarios. To facilitate a concise comparison of typical methods used in infrared small target detection,

Table 1. Performance comparison with advanced algorithms based on the HIT-UAV dataset.

Model	Size	Dataset	Precision	mAP50 (%)	Recall
YOLOv3-tiny	640	HIT-UAV	0.857	87.0 ± 0.3	0.804
YOLOv6n	640	HIT-UAV	0.889	91.4 ± 0.3	0.863
YOLOv8n	640	HIT-UAV	0.898	93.2 ± 0.1	0.9
YOLOv11 + BiFPN	640	HIT-UAV	0.912	94.2 ± 0.1	0.9
YOLOv11 + HSPAN	640	HIT-UAV	0.92	94.4 ± 0.2	0.887
YOLOv11 + CBAM	640	HIT-UAV	0.899	94.1 ± 0.2	0.902
YOLOv11 + SimAM	640	HIT-UAV	0.907	93.5 ± 0.4	0.889
YOLOv11	640	HIT-UAV	0.905	94.1 ± 0.3	0.89
YOLO-SSFA	640	HIT-UAV	0.913	94.9 ± 0.1	0.897

summarizes representative architectures and attention modules, along with their respective performance trade-offs.

3. Method

3.1. YOLO-SSFA

The proposed architecture employs YOLOv11n as the base backbone due to its lightweight design and favorable speed–accuracy trade-off. YOLO-SSFA is a lightweight model for infrared small object detection. The model optimizes the backbone network for enhanced multi-scale feature extraction by integrating convolutional neural networks and NSNs. It also includes a novel fusion mechanism in the neck structure for better feature fusion and LiteShiftHead. This lightweight detection head reduces computational complexity while improving small target detection accuracy. Figure 1 shows the overall model structure.

3.1.1. Feature Fusion Structure

Traditional object detection networks have limitations. For example, upsampling can lead to detail loss in high-resolution feature maps, and simple concatenation or weighted fusion methods struggle to capture contextual information across layers, resulting in insufficient feature representation for small targets. To address these issues, we propose an SSFF module. It uses a sequential fusion strategy to dynamically integrate low-, mid-, and high-level features via an adaptive mechanism. Depth-wise separable convolutions and attention mechanisms further enhance fusion effectiveness. The structure is shown in Figure 2, with the workflow consisting of three steps: First, global enhancement aligns low-resolution features L with mid-resolution features M using adaptive pooling, strengthening global context. Second, nearest neighbor interpolation upsamples high-resolution features to match M’s resolution, preserving details better than bilinear interpolation. Finally, aligned features L, M, and S are concatenated along the channel axis to generate a fused feature map LMS, which is input to the detection head. This enables efficient integration of multi-scale features, thereby producing richer semantic representations for accurate detection.

The proposed SSFF can be viewed as an efficiency-oriented variant of BiFPN [24]. While BiFPN adopts iterative bidirectional paths and learnable fusion weights, it incurs additional computational overhead due to repeated refinement. In contrast, SSFF performs scale-sequential fusion in a single pass, omitting complex attention layers. This simplification reduces the number of floating-point operations (FLOPs) by approximately 8.7% compared to BiFPN under the same input resolution and feature depth, as shown in

Table 2. Performance comparison with advanced algorithms based on the HIT-UAV dataset.

Model	GFLOPs	Parameters	Precision	Recall	mAP50 (%)	mAP50-95 (%)
YOLOv11	6.3	2.58	0.905	0.89	94.1 ± 0.3	59.9
YOLOv11 + BiFPN	6.3	2	0.912	0.9	94.2 ± 0.1	59.6
YOLOv11 + SSFF	5.8	2.94	0.906	0.887	94.4 ± 0.2	59.4
YOLO-SSFA	8.1	3.3	0.913	0.897	94.9 ± 0.1	60.1

. Thus, SSFF is more suitable for real-time applications on resource-limited platforms, such as UAVs and edge devices.

3.1.2. Noise Suppression Network

To improve model performance, YOLO-SSFA designed a Noise Suppression Network (NSN). NSN is a lightweight attention module that enhances key feature capture by interactively integrating channel and spatial features. Unlike traditional modules, NSN uses feature grouping and shuffling to reduce computational complexity and parameter count while leveraging inter-group correlations, making it ideal for resource-constrained scenarios. Specifically, NSN groups channels into sub-features, processes them in parallel, models dependencies using a Shuffle Unit, aggregates the sub-features, and applies channel shuffling for cross-sub-feature communication. The NSN architecture is shown in Figure 3.

The following is the detailed processing flow of the NSN module:

• Input Features. Let $Y\in {ℝ}^{B\times C\times H\times W}$ represent the input feature map, where B, C, H, and $W$ correspond to the batch size, channel count, height, and width, respectively.
• Feature grouping and channel segmentation. The feature map $Y$ is reorganized into G groups, denoted as $Y=[Y_1,\cdots ,Y_G],Y_k\in {ℝ}^{C/G\times H\times W}$, where each sub-feature $Y_k$ learns distinct semantic representations during training. An attention unit is then applied to each sub-feature to compute importance weights. Specifically, each $Y_k$ is split into two branches along the channel dimension, denoted as $Y_{k1},Y_{k2}\in {ℝ}^{C/2G\times H\times W}$.
• Channel attention calculation. The final output is computed as

  $$Y_{k1}^{\prime }=\sigma \left(F_c\left(S\right)\right)\cdot Y_{k1}=\sigma \left(W_{1s}+b_1\right)\cdot Y_{k1}$$

  (1)

Here, $W_1\in {ℝ}^{C/2G\times 1\times 1}$ and $b_1\in {ℝ}^{C/2G\times 1\times 1}$ denote the scaling and shifting parameters applied within the functions.

Spatial attention computation. The final output is

$$Y_{k2}^{\prime }=\sigma \left(W_2\cdot \mathit{GN}\left(Y_{k2}\right)+b_2\right)\cdot Y_{k2}$$

(2)

Here, $W_2$ and $b_2$ are parameters with a shape of ${ℝ}^{C/2G\cdot 1\cdot 1}$. The outputs of the two branches are merged to reestablish the original channel dimensionality.

4.

Fusion. Finally, a “channel shuffling” strategy is applied to enhance inter-group feature interaction along the channel axis. The resulting output preserves the same shape as the input $Y$, allowing seamless integration of the NSN module into modern detection architectures.

The NSN module shares the idea of spatial and channel-wise attention with CBAM [12] but introduces group-wise attention blending and a lightweight channel shuffle mechanism. This design allows efficient feature modulation while maintaining low parameter complexity, enhancing applicability on edge devices.

3.1.3. LiteShiftHead Detection Head

To overcome the balance challenge between detection precision and efficiency, we introduce LiteShiftHead, a lightweight detection head. It achieves precise and efficient target detection through feature fusion and target box regression. The overall structure is shown in Figure 4, with key components including convolution layer (Conv), regression layer (REG), classification layer (CLS), sparse convolution (SPConv), and distributed focal loss (DFL). The SPConv module structure is shown in Figure 5.

The operational mechanism of the LiteShiftHead module is structured as follows: The input feature map undergoes processing via a cascade of 3 × 3 and 1 × 1 convolutional layers, enabling the capture of multi-scale feature representations. These layers reduce dimensionality and extract semantic information for bounding box regression and object classification. Next, LiteShiftHead fuses multi-scale features using the REG module. After convolutional operations, the SPConv module optimizes each feature map by reducing parameters and enhancing efficiency. SPConv uses varying kernel sizes to extract multi-scale information, improving adaptability to targets of different sizes. The REG module predicts bounding box coordinates, ensuring detection accuracy, while the CLS module predicts object classes to enhance classification accuracy. Finally, DFL optimizes the distribution of bounding box predictions, improving localization precision. The LiteShiftHead module attains superior detection performance while minimizing computational burden, making it highly adaptable for deployment on resource-constrained platforms, including mobile devices, UAVs, and autonomous driving systems.

LiteShiftHead adopts a streamlined detection head inspired by NanoDet [17] but introduces a modular SPConv layer and DFL optimization, resulting in higher localization accuracy under constrained resources.

3.2. Datasets

This study leverages two publicly accessible datasets, HIT-UAV and FLIR, for model validation and performance evaluation. To improve the model’s generalization performance, the categories “Other Vehicles” and “DontCare” were excluded from the dataset [43]. Only the categories “Person”, “Car”, and “Bicycle” were retained. The resulting dataset comprises 2898 infrared images, each featuring a resolution of 640 × 512 pixels. Stratified sampling was employed to partition the dataset into training (2008 images), testing (571 images), and validation (287 images) sets at a ratio of 7:2:1. Additionally, the FLIR-provided infrared scene dataset was utilized for comparative and ablation experiments. This dataset consists of 10,548 infrared images, each featuring a resolution of 640 × 512 pixels, and includes three labels categorized as “pedestrian”, “bicycle”, and “car.” The same partitioning strategy was applied, dividing the dataset into training (7381 images), testing (2111 images), and validation (1056 images) sets in a 7:2:1 ratio.

3.3. Data Augmentation

To increase dataset diversity and improve model generalization, this study adopts data augmentation strategies by applying two specific augmentation schemes to the HIT-UAV dataset. For random rotation augmentation, images are rotated by +90°, −90° or remain in their original orientation with equal probability (p = 1/3). The random cropping approach simulates viewpoint variations by applying random shifts of ±10° along both the horizontal and vertical axes. After applying the above augmentation methods, the total dataset size increases to 5649 images. Specifically, the training, testing, and validation subsets are expanded to 5223, 112, and 314 images, respectively, while maintaining the original 7:2:1 partitioning ratio. The resolution of the images remains unchanged at 640 × 512 pixels, and the original three categories, “Person”, “Bicycle” and “Car” are retained.

3.4. The Evaluation Criteria

To tackle the distinct challenges of infrared small target detection—including low target–background contrast, cluttered thermal noise, and real-time inference requirements on edge devices—YOLO-SSFA is evaluated using a combination of accuracy, efficiency, and robustness metrics, with specific emphasis on small target representation and computational feasibility for UAV-based deployment. Mean Average Precision (mAP50) functions as the core metric to quantify the model’s localization accuracy for infrared small targets (particularly those ≤ 32 pixels, accounting for 70% of the HIT-UAV dataset). Precision and Recall are used to analyze false positives and false negatives in complex backgrounds. The F1 score, as the harmonic mean of Precision and Recall, reflects the model’s stability in mixed-scale infrared scenarios. Parameter count and floating-point operations (FLOPs) directly reflect the model’s lightweight design. The following equations define the evaluation metrics:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(3)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(4)

$$F1=2\ast {\displaystyle \frac{P\ast R}{P+R}}$$

(5)

$$mAP={\displaystyle \frac{1}{c}}{\displaystyle \sum _{j=1}^c{AP}_i}$$

(6)

where TP refers to true positive instances, FN is false negative, and FP stands for false positive.

Additionally, frames per second (FPS) measures the processing throughput of the model, reflecting both detection speed and real-time recognition capability. The formula for calculating FPS is given by

$$FPS={\displaystyle \frac{N}{T}}$$

(7)

where N denotes the total number of processed images, and T denotes the overall processing time.

4. Results of the Experiments

4.1. Experiment Environment

To validate the efficacy of the YOLO-SSFA framework, a series of comparative tests and ablation analyses were conducted. The experimental hardware configuration is detailed in

Table 3. Experimental platform and parameter configuration.

Names	Related Configurations
Graphics processing unit	GeForce RTX 4060 Ti (NVIDIA Santa Clara, CA, USA)
Central processing unit	Intel(R) Core i7-13700KF (Intel Santa Clara, CA, USA)
GPU memory size	8 G
Operating system	Win 11
Computing platform	CUDA12.6
Deep learning framework	Pytorch2.1.0
Development Environment	Pycharm 2024.1.1 x64
Image pixels	640 × 640
Batch size	16
Epochs	300

. The model was trained for 300 epochs, with a batch size of 16 and an input resolution of 640 × 640 pixels. The SGD optimizer was used with an initial learning rate of 0.01, momentum of 0.937, and a weight decay of 5 × 10⁻⁴. No dropout was applied. A fixed learning rate schedule was used without cosine decay or warm restarts. The loss functions consisted of Complete IoU (CIoU) loss for bounding box regression, Binary Cross-Entropy (BCE) loss for classification, and Distribution Focal Loss (DFL) for improved localization accuracy. Early stopping was enabled with a patience of 100 epochs, and model checkpoints were selected based on the best validation mean Average Precision (mAP). Automatic Mixed Precision (AMP) training was enabled to accelerate convergence. Most importantly, all training did not use pretrained weights. Each model results are averaged over three independent runs with different random seeds. The reported performance includes mean ± standard deviation.

4.2. Comparison Studies and Analysis

To comprehensively evaluate the applicability and effectiveness of YOLO-SSFA in UAV-based infrared target detection, extensive comparisons were made against multiple state-of-the-art algorithms using the HIT-UAV and FLIR datasets under consistent experimental settings.

As shown in

Table 4. Performance comparison with advanced algorithms based on the HIT-UAV dataset.

Model	Size	Parameters	GFLOPs	F1 (%)	Precision	mAP50 (%)	mAP50-95 (%)	FPS
YOLOv3-tiny	640	12.13	18.9	82	0.857	87.0 ± 0.3	52	146.1
YOLOv5n	640	2.5	7.1	90	0.888	93.4 ± 0.2	57.9	158.3
YOLOv6n	640	4.23	11.8	88	0.889	91.4 ± 0.3	56.1	179.5
YOLOv8n	640	3.01	8.1	90	0.898	93.2 ± 0.1	58.9	187.5
YOLOv8s	640	11.13	28.4	90	0.909	94.2 ± 0.1	61.3	122.5
YOLOv9t	640	2	7.8	90	0.891	93.9 ± 0.3	59.4	154
YOLOv10n	640	2.7	8.2	89	0.893	93.4 ± 0.2	59.3	191.6
YOLOv11n	640	2.58	6.3	90	0.905	94.1 ± 0.3	59.5	165.5
YOLO-SSFA	640	3.3	8.1	91	0.913	94.9 ± 0.1	60.1	217.6

, while YOLOv8s achieved the second-highest mAP50 of 94.2%, the YOLO-SSFA model demonstrated superior real-time capability, with a detection speed of 217.6 FPS, which is 34.6% faster than YOLOv8s (122.5 FPS) and 31.5% faster than the baseline YOLOv11n (165.5 FPS). Notably, YOLO-SSFA achieved the highest mAP50 of 94.9%, outperforming YOLOv11n by 0.8% and YOLOv5n by 1.5% while maintaining lightweight characteristics with only 3.3 M parameters—a 28% increase over YOLOv11n (2.58 M) but with 70% fewer parameters than YOLOv8s (11.13 M). In terms of computational efficiency, YOLO-SSFA’s GFLOPs (8.1) are comparable to YOLOv8n (8.1) but achieve 0.7% higher mAP50. Compared to heavier models like YOLOv3-tiny (18.9 GFLOPs), YOLO-SSFA reduces computational complexity by 57% while improving mAP50 by 8.9%. The model also outperforms YOLOv9t (93.9% mAP50) by 1.0% in accuracy and 41.3% in FPS, demonstrating a superior balance of Precision and speed. Notably, YOLO-SSFA’s F1 score (91%) and Precision (0.913) are the highest among all tested models, indicating strong generalization ability in infrared small-target scenarios. Its lightweight architecture (3.3 M parameters, 8.1 GFLOPs) enables a 217.6 FPS inference speed, making it 1.3× faster than YOLOv10n (191.6 FPS) and 1.5× faster than YOLOv5n (158.3 FPS) while maintaining competitive detection accuracy. Figure 6 shows the visualization results and heatmaps on the HIT-UAV dataset.

To further evaluate the detection performance of the proposed YOLO-SSFA model, we plotted the Precision–Recall (P–R) curves for both the HIT-UAV and FLIR datasets, comparing them with the baseline YOLOv11n. As shown in Figure 7, YOLO-SSFA consistently exhibits superior Precision and Recall trade-offs across both datasets.

These curves demonstrate that the integration of SSFF, NSN, and LiteShiftHead modules not only improves overall accuracy but also enhances stability across varying detection thresholds.

The performance comparison on the FLIR dataset (dominated by complex backgrounds and medium–small targets) in

Table 5. Performance comparison with advanced algorithms based on the FLIR dataset.

Model	Size	Parameters	GFLOPs	F1 (%)	Precision	mAP50 (%)	mAP50-95 (%)	FPS
YOLOv3-tiny	640	12.13	18.9	73	0.806	73.7 ± 0.3	42.2	281.1
YOLOv5n	640	2.5	7.1	80	0.841	83.6 ± 0.1	49.2	384.0
YOLOv6n	640	4.23	11.8	78	0.819	81.5 ± 0.2	48.2	381.7
YOLOv8n	640	3.01	8.1	80	0.858	83.2 ± 0.1	49.8	332.1
YOLOv8s	640	11.13	28.4	82	0.85	85.7 ± 0.1	52.9	189.4
YOLOv9t	640	2	7.8	79	0.846	83.5 ± 0.3	49.9	292.9
YOLOv10n	640	2.7	8.2	78	0.829	82.5 ± 0.2	49.3	355.4
YOLOv11n	640	2.58	6.3	79	0.85	83.8 ± 0.3	50.8	299
YOLO-SSFA	640	3.3	8.1	80	0.867	85 ± 0.2	51.4	319.4

highlights YOLO-SSFA’s robustness and efficiency. With an mAP50 of 85.0%, YOLO-SSFA outperforms YOLOv11n (83.8%) by 1.2% and YOLOv5n (83.6%) by 1.4% while achieving a detection speed of 319.4 FPS—6.8% faster than YOLOv11n (299 FPS) and 68.7% faster than YOLOv8s (189.4 FPS). The model’s lightweight architecture (3.3 M parameters, 8.1 GFLOPs) ensures low computational overhead, with a parameter count 70% lower than YOLOv8s (11.13 M) and GFLOPs 71.5% lower than YOLOv3-tiny (18.9 GFLOPs). Notably, YOLO-SSFA’s F1 score (80%) and Precision (0.867) surpass YOLOv6n (78%, 0.819) and YOLOv10n (78%, 0.829), demonstrating stronger target–background discrimination in cluttered infrared scenes. The NSN module plays a critical role here, reducing false alarms by suppressing background noise, while the LiteShiftHead maintains high throughput via sparse convolution. Compared to the HIT-UAV dataset, YOLO-SSFA’s mAP50 decreases slightly (−0.9%) on FLIR but achieves a 15.5% higher FPS, reflecting its adaptability to mixed-scale target scenarios. Figure 8 shows the visualization results and heatmaps on the HIT-UAV dataset.

Figure 9 shows the Precision–Recall curve of YOLOv11n and YOLO-SSFA on the FLIR dataset.

To further evaluate the performance and generalization of YOLO-SSFA, we conducted comparative experiments with transformer-based and anchor-free detectors on both the HIT-UAV and FLIR datasets. As shown in

Table 6. Performance comparison with advanced algorithms based on the HIT-UAV dataset.

Model	GFLOPs	Parameters	mAP50 (%)	mAP50-95 (%)		AP50 (%)		FPS
					Bicycle	Car	Person
YOLOv11n	6.3	2.58	94.1 ± 0.3	59.9	92.8	97.9	91.5	165.5
DETR	60.53	41.56	84.8 ± 0.1	45.4	50.5	75.4	32.6	51.2
Swin Transformer	172	36.86	72.5 ± 0.2	33.5	41.5	53.7	24.0	51.9
YOLOX	5.6	2.03	94.1 ± 0.2	57.8	92.3	97.7	91.1	225.9
YOLO-SSFA	8.1	3.3	94.9 ± 0.1	60.1	93.4	98.7	92.6	217.6

(HIT-UAV), we included DETR [39], Swin Transformer [40], and YOLOX [41] as baselines. All models were trained from scratch under the following identical conditions: input resolution of 640 × 640 pixels, batch size 16, and 300 epochs.

As shown in

Table 7. Performance comparison with advanced algorithms based on the FLIR dataset.

Model	GFLOPs	Parameters	mAP50 (%)	mAP50-95 (%)		AP50 (%)		FPS
					Bicycle	Car	Person
YOLOv11n	6.3	2.58	83.8 ± 0.3	50.8	78.8	92.2	82.4	299
DETR	60.53	41.56	81.6 ± 0.1	44.2	56.8	65.6	35.5	51.4
Swin Transformer	172	36.86	76.6 ± 0.1	41.0	53.6	61.3	32.8	50.5
YOLOX	5.6	2.03	84.7 ± 0.1	48	78.6	92.3	82.1	288.54
YOLO-SSFA	8.1	3.3	85 ± 0.2	51.4	80.1	92.4	83.1	319.4

(FLIR), these results across both datasets confirm that YOLO-SSFA not only achieves competitive accuracy compared to recent transformer and anchor-free models, but also surpasses them in inference speed and efficiency, demonstrating its suitability for real-time, resource-constrained infrared target detection scenarios.

As depicted in Figure 10, this study conducts a comprehensive evaluation of three critical metrics, mAP50, FPS, and model size, from the standpoint of model practicality. These metrics correspond to detection accuracy, inference speed, and resource efficiency, respectively. The experimental findings indicate that YOLO-SSFA attains superior detection accuracy and inference speed while preserving a compact architectural design. This outcome substantiates that YOLO-SSFA strikes an optimal balance between accuracy and speed, with its lightweight structure providing substantial benefits for mobile deployment. In summary, YOLO-SSFA exhibits exceptional overall performance in infrared target detection, positioning it as a robust and practical solution for real-world deployments.

4.3. Ablation Studies and Analysis

To assess the impact of the proposed YOLO-SSFA model, extensive ablation tests were performed using both the HIT-UAV and FLIR datasets. The objective was to analyze how each individual component contributes to the model’s overall performance. The corresponding experimental results are summarized in

Table 8. Ablation analysis conducted on the HIT-UAV dataset.

Yolov11n	SSFF	LiteShiftHead	NSN	Parameters	F1 (%)	Precision	mAP50 (%)	mAP50-95 (%)	FPS
√				2.58	90	0.905	94.1 ± 0.3	59.9	165.5
√	√			2.94	89	0.906	94.4 ± 0.2	59.4	144.4
√		√		2.60	90	0.908	94.2 ± 0.2	59.9	189.4
√			√	2.83	89	0.898	94.3 ± 0.1	59.6	220.5
√	√	√		3.01	90	0.911	94.6 ± 0.2	60.0	155.8
√	√	√	√	3.3	91	0.913	94.9 ± 0.1	60.1	217.6

. On the HIT-UAV dataset, which predominantly contains small targets, the SSFF module alone improves mAP50 from 94.1% (baseline YOLOv11n) to 94.4%, validating its effectiveness in enhancing multi-scale feature representation for tiny objects. Adding the NSN module boosts FPS by 33.3% (from 165.5 to 220.5), indicating its role in optimizing computational efficiency without significant accuracy loss. Integrating SSFF, LiteShiftHead, and NSN achieves the highest mAP50 of 94.9% (+0.8% over the baseline), with a parameter increase of only 27.9%, demonstrating the synergistic effect of multi-module design for small-target scenarios.

On the FLIR dataset, characterized by complex backgrounds,

Table 9. Ablation analysis conducted on the FLIR dataset.

Yolov11n	SSFF	LiteShiftHead	NSN	Parameters	F1 (%)	Precision	mAP50 (%)	mAP50-95 (%)	FPS
√				2.58	79	0.85	83.8 ± 0.3	50.8	299
√	√			2.94	79	0.841	84.3 ± 0.1	50.9	285.7
√		√		2.60	79	0.858	84.4 ± 0.2	51.0	312.2
√			√	2.83	80	0.844	84.1 ± 0.1	50.8	327.6
√	√	√		3.01	80	0.86	84.6 ± 0.2	51.2	290.9
√	√	√	√	3.3	80	0.867	85 ± 0.2	51.4	319.4

indicates that the LiteShiftHead alone improves FPS by 7.8% (from 299 to 312.2), with a mAP50 gain (+0.6%), highlighting its efficiency in inference acceleration. The NSN module significantly enhances F1 score from 79% to 80% and Precision from 0.85 to 0.844, likely due to its ability to suppress background noise in cluttered scenes. When all modules are integrated, YOLO-SSFA maintains 85.0% mAP50 (comparable to YOLOv8s) while achieving 319.4 FPS (+6.8% over the baseline), proving the robustness of the lightweight architecture across diverse infrared scenarios.

Figure 11 demonstrates the enhanced performance of the YOLO-SSFA model over the original algorithm in detecting infrared vehicle targets. Consequently, the YOLO-SSFA model represents a more suitable and effective solution for detecting weak small targets in infrared environments.

4.4. Comparative Analysis of Various Attention Mechanisms

To further verify the validity of the proposed YOLO-SSFA model, a comparative experiment on attention mechanisms was conducted using the HIT-UAV dataset. This experiment included several widely used attention mechanisms, and only the attention mechanism was modified during the experiment.

Table 10. Comparative analysis of attention mechanisms conducted on the HIT-UAV dataset.

Yolov11n	SSFF	LiteShiftHead	Attention	Parameters	F1 (%)	Precision	mAP50 (%)	mAP50-95 (%)	FPS
√	√	√	CBAM	3.36	90	0.91	94.2 ± 0.2	59.7	182.7
√	√	√	GAM	4.2	90	0.912	93.6 ± 0.4	58.4	182.0
√	√	√	SimAM	3.34	90	0.901	93.9 ± 0.3	58.8	164.7
√	√	√	SK	14.4	90	0.909	94.5 ± 0.3	59.8	147.3
√	√	√	EMA	3.5	91	0.915	93.9 ± 0.1	59.9	165.3
√	√	√	NSN	3.3	91	0.913	94.9 ± 0.1	60.1	217.6

presents a performance comparison of the YOLO-SSFA model integrated with different attention mechanisms on the HIT-UAV dataset, highlighting the NSN module’s superior lightweight design and detection efficiency. With an mAP50 of 94.9% and a detection speed of 217.6 FPS, NSN outperforms all comparative schemes; its mAP50 exceeds CBAM [12] (94.2%), GAM [44] (93.6%), SimAM [13] (93.9%), SK [45] (94.5%), and EMA [46] (93.9%) by 0.7%, 1.3%, 1.0%, 0.4%, and 1.0%, respectively. Meanwhile, NSN achieves a 47.7% increase in FPS compared to SK (147.3 FPS) and a 31.6% increase compared to EMA (165.3 FPS), demonstrating significant real-time performance advantages. In terms of lightweight design, NSN introduces only 0.72 M parameters (total parameters: 3.3 M), reducing the parameter count by 77% compared to SK (14.4 M) and 21% compared to GAM (4.2 M) while achieving a 1.3% mAP50 improvement over GAM. By partitioning channels into subgroups and introducing Shuffle Units, NSN effectively suppresses background noise while avoiding the computational redundancy of traditional attention modules (e.g., CBAM’s sequential convolutions). Experimental data show that NSN achieves the highest F1 score (91%) and Precision (0.913), indicating superior discrimination between small targets and cluttered backgrounds, especially in low-contrast scenarios. In summary, NSN achieves an optimal balance of accuracy, speed, and parameter efficiency through channel-group sparse computation and dynamic feature shuffling. These results validate NSN’s effectiveness in enhancing feature representation and suppressing noise for infrared small-target detection, making it an ideal solution for real-time deployment on edge devices.

In addition, to verify the effectiveness of the NSN module, we used Grad CAM to visualize the spatial attention map of YOLO-SSFA. Figure 12 and Figure 13 show the comparison of attention response between our model and YOLO-SSFA without attention mechanism.

The results demonstrate that YOLO-SSFA exhibits more concentrated and target-focused activation regions. In particular, small objects such as distant pedestrians or bicycles generate clear, localized responses, while background clutter is largely suppressed. This suggests that the integration of channel–spatial grouping and channel shuffle in NSN enhances the network’s ability to focus on semantically meaningful areas, contributing to improved precision.

5. Conclusions

This study presents the YOLO-SSFA model, which enhances infrared small target detection through multi-module optimization. It achieves mAP50 scores of 94.9% and 85% on the HIT-UAV and FLIR datasets, respectively. By integrating three key modules, SSFF, LiteShiftHead, and NSN, the model enhances detection accuracy and robustness. The SSFF module strengthens feature extraction for weak and small targets via multi-scale fusion. The LiteShiftHead module reduces computational cost while maintaining accuracy and accelerating inference through sparse convolution. The NSN module focuses on salient features and suppresses background noise using attention mechanisms. Experimental results indicate that the YOLO-SSFA model surpasses traditional methods in both detection accuracy and computational efficiency, demonstrating strong potential for real-world deployment.

Although the proposed YOLO-SSFA model demonstrates high inference speed and accuracy in experimental evaluations, it is important to note that the current testing setup does not fully replicate a real-world real-time deployment scenario, such as on embedded systems or UAVs with limited computational resources. To implement the model in an actual real-time environment, further optimizations would be necessary, including quantization, pruning, and hardware-specific deployment. Additionally, real-time data acquisition, thermal noise adaptation, and latency control in embedded pipelines should be addressed in future work.

The proposed method has potential applications in both civilian and security-related domains. All datasets used in this study are publicly available (HIT-UAV and FLIR), and no sensitive or military data were involved. We fully comply with ethical guidelines and data-sharing policies as outlined by the respective dataset providers and journal requirements. The code is available at the following URL: https://github.com/dd001217/YOLO-SSFA.git, accessed on 1 June 2025.