AFR-Net: Attention-Based Channel-Spatial Feature Refinement Network for Small Object Detection

Small object detection (SOD) is a critical yet challenging area in computer vision. Deep learning-based object detection techniques have achieved remarkable progress, however, identifying small objects continues to be a challenging task due to factors such as scale variations, object deformation, and interference from complex backgrounds. The Feature Pyramid Network (FPN) partially mitigates these challenges and significantly enhances detection performance, but existing FPN-based approaches often face substantial information loss at different levels of the pyramid, limiting the effectiveness of multi-scale feature utilization. To address these challenges, we propose AFR-Net to improve small object detection by augmenting the functionality of feature pyramid networks. AFR-Net comprises three primary modules. Multi-scale Deformable Feature Extraction (MDFE) leverages deformable convolution to enhance feature extraction at the deepest FPN level, generating more accurate feature maps. During the feature fusion process, Channel-Spatial Feature Refinement (CSFR) module employs attention mechanisms to refine important feature information. Wise Intersection over Union (WIoU) emphasizes small objects by assigning appropriate weights to different scales. We evaluate AFR-Net on benchmark datasets, including MS COCO and VisDrone, demonstrating that it achieves superior performance compared to existing FPN-based detection methods. Furthermore, AFR-Net’s design makes it applicable to various feature-refinement-based tasks in computer vision, enabling improved performance across multiple domains.