Enhanced YOLOv7 for Lightweight Drone Object Detection

A lightweight object detection algorithm based on YOLOv7 is proposed, optimized for deployment on resource-constrained drone platforms. Key improvements include optimizing the network structure, adjusting depth and channel dimensions, and integrating the Bi-directional Feature Pyramid Network (BiFPN) to enhance multi-scale feature fusion for improved detection accuracy and robustness. A 2-Level Haar Wavelet Transform (2-level HWT) replaces traditional strided convolutions, efficiently preserving spatial information for better feature extraction. Additionally, a novel Dual Cross Stage Partial Fusion (DCSPF) module is introduced to improve feature reuse and representation capacity by facilitating information flow across stages. The detection head is further enhanced with the Group Batch Normalization Reweighting Unit (GNRU) and the Adaptive Grouped Convolution Unit (AGCU) to dynamically adjust feature importance. Experimental results on the Vis-Drone2019 dataset show a 2.6% increase in mAP@50 and a 6.1% reduction in GFLOPs, validating the proposed approach's balance between accuracy and computational efficiency, suitable for drones with limited resources.