Object Detection in YOLOv5 Remote Sensing Images Based on Reparameterization

An improved object detection algorithm based on Yolov5s is proposed for the characteristics of remote sensing images with many types of features, dense objects and complex background. Firstly, the idea of reparameterization is applied in the backbone network to reduce the computational amount of network detection; secondly, the Efficient Channel Attention (ECA) mechanism is applied in the Yolov5s backbone network, which is used to enhance the feature extraction ability of the network; finally, the second layer of the backbone network is replaced by the Dilated convolution with a larger perceptual field under the same number of parameters. The experimental results show that the algorithm in this paper achieves an accuracy of 72.90% on the DOTA dataset, which is 1.76% higher than the mean accuracy (mAP) of Yolov5s, and the computational GFLOPS ranges from 17.4 to 17.5, which is a large improvement in accuracy obtained with only a small number of parameter additions compared to Yolov5s.