Lightweight Small Target Detection Algorithm Based on Feature Enhancement and Channel Weight Fusion

This paper proposes a lightweight small object detection algorithm based on feature enhancement and channel weight fusion, termed FECWF-YOLO (Feature Enhancement and Channel Weight Fusion YOLO), aiming to address the challenges of insufficient detection accuracy for small objects and inadequate model lightweighting in scenarios such as unmanned aerial vehicles. The paper first introduces a small object detection head to enhance the model’s capability in capturing small objects, optimizes the network structure to improve detection performance, and reduces the number of convolutional layers in the backbone network to achieve lightweighting. To mitigate the increased computational cost incurred by the small object detection head, a lightweight detection head (LDH) is designed using the Partial Convolution (PConv) module, effectively reducing computational complexity. Furthermore, a lightweight feature enhancement module based on PConv is embedded in the backbone network, utilizing multi-branch dilated convolutions to expand the receptive field and enhance feature extraction for small objects. In the feature fusion stage, a channel weight fusion-based feature fusion method is proposed, incorporating Deformable Convolutional Networks v2 (DCNv2) to improve the effectiveness of multi-feature fusion. Experimental results show that, compared with the baseline model YOLOv8n, FECWF-YOLO reduces the parameter count and computational complexity by 70% and 9%, respectively, while achieving 5.3 percentage points increase in mAP@0.5. These results indicate a favorable balance between performance and efficiency, providing an effective solution for small object detection in scenarios such as unmanned aerial vehicles.