YOLOv8-AMCD: Improved YOLOv8 for Small Object Detection

Aiming at the problem of partially missing detection and false detection, large amount of calculation and parameters in small object detection tasks in motion scenes. This paper proposes a YOLOv8-AMCD model to improve the model's accuracy while keeping its lightweight design. Firstly, the original downsampling structure of YOLOv8 is replaced with the advanced ADown structure, which reduced the number of model parameters. Secondly, Mixed Local Channel Attention is integrated before the detection head of each layer to help the model obtain the feature information of different levels. Thirdly, part of C2f Block in the backbone network are replaced with ConvNextv2 Block to increase channel competition. Finally, Deformable Attention is added to the last layer of the backbone network, so as to obtain more effective target feature information before feature fusion. Experimental results on the public datasets VOC, DOTAv2, and the self-built dataset TableTennis show that the proposed algorithm has higher precision and mAP than the baseline model YOLOv8n, but also lower the number of parameters and the GFIOPs.