Wreckage Target Recognition in Side-scan Sonar Images Based on an Improved Faster R-CNN Model
Yulin Tang, Jin Shaohua, Bian Gang, Zhang Yonzhou, Li Fan
Abstract
Conventional object recognition approaches suffer from a number of problems such as difficulty in feature design, low detection accuracy and reliability, and weak generalization ability. This paper addresses these issues by proposing an improved Faster R-CNN model based on the VGG-16 convolutional neural network (CNN) for the automatic recognition of wreckage targets in side-scan sonar images. The object classification results of the Faster R-CNN model are improved by equalizing the number of anchor boxes in the region proposal network that either contain or do not contain wreckage targets and employing a balanced sampling of the image database for model training. The feasibility of the proposed model is demonstrated experimentally, and the results show that the average wreckage detection accuracy of the improved model is increased by 4.30% to 87.72% relative to that of the conventional Faster R-CNN model, while providing similar detection efficiency.