基于多阶段特征交互的目标检测方法OA

To address the challenges of small object density and occlusion in object detection in complex combat environments,an object detection method based on multi-stage feature interaction is proposed.First,a super-span multi-convolution module is constructed in the backbone network to enhance the feature representation capability of the model.Second,Efficient RepGFPN is introduced in the neck network to extract more discriminative fine-grained features.Finally,the EIoU loss function is used to reduce the deviation between the object box and the predicted box.Experiments on public datasets show that the mAP performance is improved by 4.4%and 4.1%respectively compared with the original YOLOv5s model,with an average processing speed of 75FPS.