Research on Lightweight Remote Sensing Image Object Detection Algorithm

Abstract

In order to solve the problems of low accuracy and slow speed that appear in traditional remote sensing detection algorithms, this paper proposes a remote sensing image object detection algorithm yolov5sf based on yolov5s improvement. The more lightweight ghost convolution module is used in the yolov5sf algorithm to reduce the redundancy in the feature map, and the hard-swish activation function is used to avoid a large number of exponential operations. We embed the GC attention module in the csp block to capture the dependencies between different channels and suppress useless features. Meanwhile, to address the problem that the network is not accurate for oversized targets detection, an mixed receptive field feature enhancement module is proposed to be placed in the downsampling layer for adaptive scaling of features of different receptive fields. The Yolov5sf algorithm is lightweight enough, with only 4.4M parameters, and only 65% of the computational effort of yolov5s and 6.7% of yolov3. Our map-50 metrics on the RSOD dataset was reduced by only 0.4%, with a 1.2% improvement in detection accuracy for the oversize target overpass category.