| Popis: |
An improved UAVAI-YOLO model was proposed to address the problem of poor target detection in UAV aerial images. Firstly, in order to obtain richer semantic information for the model, the original convolution of the C2f module of the original backbone part was replaced with the improved DCN convolution. Secondly, in order to increase the P2 feature layer without increasing the number of model parameters, the Conv_C module was proposed to downscale the output channel of the backbone network, and at the same time, because of avoiding the loss of semantic information due to channel downsizing, the original convolution of the C2f module in the neck part was replaced by the improved ODConv dynamic convolution. Then, the BIFPN module was introduced to make full use of the contextual semantic information. Finally, Wise-IoU was used to replace the original loss function to improve the accuracy of the model target detection frame. Experimental results on the publicly available VisDrone2019 dataset and UAVDT dataset showed that the UAVAI-YOLO model improves 4.4% and 1.1% compared to the original YOLOv8n model mAP0.5, respectively, high detectability accuracy compared to other mainstream object detection models. |
