Accurate identification and positioning of small targets (such as vehicles, buildings, and vegetation) in large-scale remote sensing images are crucial for military reconnaissance, urban planning, environmental monitoring, and other fields. However, traditional target detection methods often struggle to accurately identify these targets due to their small size, irregular shape, complex background, and illumination changes in the image. Therefore, there is a critical need for specialized research on small target detection. Research in this area can enhance the accuracy and efficiency of remote sensing image analysis, providing more reliable data support for decision-making and planning across various fields. This research holds significant theoretical and practical value.

Optical remote sensing images may suffer from low target detection accuracy due to complex backgrounds, varied scales, generally small targets, and different orientations. We propose a method for remote sensing small target detection based on multi-scale information fusion. Key improvements include: 1) C3 module integration: designed to integrate a global context module, enhancing the model’s ability to distinguish targets from backgrounds. This ensures the model focuses on key areas while ignoring unnecessary ones, which effectively improves target localization accuracy. 2) Optimized PANet with BiFPN: to balance feature information across different scales and strengthen multi-scale target detection performance, we optimize the PANet and introduce the BiFPN. This feature pyramid network structure better utilizes multi-level feature information for accurate detection of targets of various sizes. 3) Circular smooth label method: addressing the challenge of targets at different directions and angles, this method transforms the true rotation angle of target objects into a continuous probability distribution. This approach converts the angle regression problem into a classification problem, thereby improving detection and positioning accuracy. 4) Image slicing preprocessing: to enable rapid detection of high-resolution images, we adopt an image slicing preprocessing method, which segments large images into smaller blocks for processing, significantly reducing false detection and missed detection of small targets.

To thoroughly validate the effectiveness of the proposed algorithm, we conduct a series of module ablation experiments on the DOTA dataset, with the experimental results detailed in Table 1. Based on the data shown in Table 1, our study successfully enhances the model’s feature extraction capabilities, which strengthens its accuracy in locating target areas and achieves an algorithmic mAP of 83.7%. To further assess the performance of the improved algorithm, we make comparisons with advanced target detection algorithms such as R²CNN, YOLOv3, SCRDet, YOLOv5s, YOLOv6s, MaskOBB and YOLOv7 using the DOTA dataset. The experimental findings are summarized in Table 2. The analysis of these results demonstrates that the algorithm proposed in this study outperforms other comparison algorithms in terms of accuracy. To comprehensively evaluate the performance of the GCB-YOLOv5 algorithm, we employ the same remote sensing dataset for verification, comparing its detection rates with those of the original YOLOv5 algorithm and other algorithms in the YOLO series. The findings are presented in Table 3.

In the face of challenges such as diverse target scales, complex backgrounds, the prevalence of small targets, and diverse target orientations in optical remote sensing images, we first introduce the GCC3 module designed to enhance the model’s ability to distinguish between targets and backgrounds. This enhancement directs the model’s focus towards key areas while disregarding unnecessary ones, thereby significantly improving the detection accuracy of small-scale targets. Additionally, our study replaces the PANet structure with BiFPN to better address the detection requirements of multi-scale targets. The incorporation of circular smooth labeling effectively manages the multi-scale and directional uncertainties of the targets. The experimental results strongly support the significant advantages of the proposed algorithm in small-scale target detection. In future research, the model will be optimized for lightweight performance to balance the reasoning speed and detection accuracy, thereby enhancing its applicability in practical scenarios.