Research on Target Detection and Recognition Algorithms in Remote Sensing Images

Abstract

With the continuous improvement of computer vision and deep learning technology, the target detection methods of remote sensing images are also expanding and diversifying. In view of the shortcomings of the current object detection and recognition algorithms in terms of accuracy and versatility, this paper introduces the reverse scale transfer layer and feature pyramid (FPN) modules and applies the attention models of channel attention mechanism and spatial attention mechanism to each module of the convolutional neural network, so that the feature layer can obtain accurate and comprehensive prediction information, and finally proposes a remote sensing image object detection algorithm DCYOLOv7 with high accuracy. Compared with the benchmark model, the accuracy of the algorithm on small, medium, and large targets is improved by 14.69%, 4.14%, and 5.19%, respectively. The DC-YOLOv7 algorithm is improved by 10.15%, 12.16%, 13.18%, and 14.8% compared with the mAP, AP50, AP75, and AR100 of the benchmark algorithm, respectively. DC-YOLOv7 has a better detection application effect than the classical algorithm in the military aspect. The effectiveness and versatility of the target detection and recognition algorithm in the remote sensing images presented in this paper have been verified.