Research on small target recognition technology based on YOLOv8

In order to solve the challenges of small objects in UAV view that are not easily recognized, complex background and severe occlusion, we improved on the original YOLOv8s algorithm, enhancing it’s ability to capture and fuse small target information. We introduce RFAConv structure, which enables model to better extract feature information at different scales by adaptively adjusting the sensing field. Second, after multiple downsampling, small target features are easily blurred or even lost, resulting in the model not being able to effectively fuse the shallow and deep information when the features are fused. For this reason, we designed Omni-Fusion FPN module, which not only can fully fuse the contextual information, but also improves Model’s ability to capture and fuse the small target information. Finally, we use to Dyhead replace detection head of the original model, which enhances the detection capability of targets at different scales by dynamically fusing multi-level contextual details. We conduct experiments on VisDrone2019 dataset, and compared to YOLOv8s model, the improved algorithm improves $1.9 \%, 4.2 \%$, $4.7 \%$, and $3.4 \%$ on Precision, Recall, mAP50, and mAP50-95, respectively. Compared to other mainstream target detection models, our improved algorithm shows better detection performance.