Integrate YOLOv3 with a Self-attention Mechanism for Underwater Object Detection Based on Forward-looking Sonar Images

Abstract

Detection of underwater objects based on sonar images is an important component of underwater robotics in the field of environmental perception and has attracted much attention. However, due to the complexity of the underwater environment, sonar devices, such as forward-looking sonar, encounter interference noises during imaging. In addition, there are few public sonar datasets that can be used for research. These challenges limit the task of underwater object recognition based on Deep Learning and make it difficult to make progress. Therefore, in this study, we first public a forward-looking sonar image dataset (FLSD) with rich data types and a large number of images. Then, we proposed a data augmentation method to address the lack of sonar images, and an experiment demonstrated its effectiveness. Finally, to strengthen the correlation of feature maps between channels in the training process of the network, we introduced the YOLOv3-based self-attention mechanism and proposed YOLOv3-SE. The results of our experiments show an improvement of 2.94% Precision, 3.34% Recall and 4.95% mAP on FLSD. Dataset attachment: https://code.ihub.org.cn/projects/14186