A lightweight enhanced YOLOv8 algorithm for detecting small objects in UAV aerial photography

Facing with the problems of tiny and densely distributed image targets, uneven class distribution, and model size restriction of hardware conditions, object detection from the perspective of drones has less precise results. To tackle this issue, we introduce a refined YOLOv8m model integrating various lightweight strategies. Firstly, we propose an ESEMB (Mobile inverted Bottleneck convolution with Effective Squeeze and Extraction) module and a Faster module to improve c2f in the backbone. Specifically, ESEMB balances the three dimensions of network depth, width and image resolution for lightweight, and Faster module reduces the calculation of redundant channels to achieve efficient feature extraction. Secondly, we optimize the neck structure by adding feature layers with abundant semantic information of tiny targets, modifying the fusion mechanism with adding additional fusion of underlying information to increase spatial semantic information of feature graph, which are conducive to small object detection. Thirdly, the improved detection head Phead notably reduces the number of model parameters by redesigning the detection head using lightweight module. Additionally, we use inner-IoU loss function to improve MPDIoU (Minimum Point Distance based IoU) function, replacing the original loss function with inner-MPDIoU, improving the learning ability of difficult samples. Finally, we use a variety of pruning algorithms to further lightweight the model and fine-tune it. The experimental results show that the model size and parameter number of the improved model are only 14.2% and 49.6% of the original YOLOv8m model, but the model detection performance has been considerably improved. P, mAP@0.5 and mAP@0.5:0.95 have respectively increased by 15.9%, 21.7% and 34.1%. It achieves lightweight and optimization accuracy at the same time, and can be effectively applied to target detection tasks on UAV platforms. The code and datasets are available at https://github.com/pwvivi/le-yolo.