Automatic Bridge Detection of SAR Images Based on Interpretable Deep Learning Algorithm

As a typical feature target in SAR images, the accurate detection and localization of bridge targets are of important research value for the management of transportation facilities and battlefield reconnaissance. Since most of the current bridge target detection algorithms use horizontal frames, they cannot achieve accurate localization of bridges. There is a large amount of background information in the data set annotation, which is not conducive to the training of the network. In this paper, an in-depth analysis of the interpretable deep learning algorithm is presented to show that the use of rotating detection frames is more beneficial to the automatic detection of targets with large aspect ratios such as bridges. The algorithm framework first performs feature extraction of bridge targets through the Yolov5 network, and then passes the feature maps containing target information into the proposed regression prediction visualization module to obtain detection result maps and heat maps. In the experiments, the SAR images of the Foshan area are used to verify the results, and the advantages of the rotating detection frame in terms of localization and inspection performance are analysed by comparing the IoU and the heat map.