文献类型：期刊文献

英文题名：A Shadow Capture Deep Neural Network for Underwater Forward-Looking Sonar Image Detection

作者：Xiao, Taowen[1];Cai, Zijian[1];Lin, Cong[1];Chen, Qiong[2]

机构：[1]Guangdong Ocean Univ, Coll Elect & Informat Engn, Zhanjiang 524025, Peoples R China;[2]Tsinghua Univ, Inst Global Change Studies, Minist Educ, Dept Earth Syst Sci,Key Lab Earth Syst Mod, Beijing 100084, Peoples R China

年份：2021

卷号：2021

外文期刊名：MOBILE INFORMATION SYSTEMS

收录：SCI-EXPANDED(收录号：WOS:000741076700001)、、EI(收录号：20220311477440)、Scopus(收录号：2-s2.0-85122862391)、WOS

基金：AcknowledgmentsTaowen Xiao, Zijian Cai, Cong Lin, and Qiong Chen contributed equally to this work. This work was supported by the National Natural Science Foundation of China under Grant 62072121 and Natural Science Foundation of Guangdong Province 2021A1515011847.

语种：英文

外文关键词：Convolutional neural networks - Object detection - Image enhancement - Neural network models - Object recognition - Sonar - Underwater acoustics

外文摘要：Image sonar is a widely used wireless communication technology for detecting underwater objects, but the detection process often leads to increased difficulty in object identification due to the lack of equipment resolution. In view of the remarkable results achieved by artificial intelligence techniques in the field of underwater wireless communication research, we propose an object detection method based on convolutional neural network (CNN) and shadow information capture to improve the object recognition and localization effect of underwater sonar images by making full use of the shadow information of the object. We design a Shadow Capture Module (SCM) that can capture the shadow information in the feature map and utilize them. SCM is compatible with CNN models that have a small increase in parameters and a certain degree of portability, and it can effectively alleviate the recognition difficulties caused by the lack of device resolution through referencing shadow features. Through extensive experiments on the underwater sonar data set provided by Pengcheng Lab, the proposed method can effectively improve the feature representation of the CNN model and enhance the difference between class and class features. Under the main evaluation standard of PASCAL VOC 2012, the proposed method improved from an average accuracy (mAP) of 69.61% to 75.73% at an IOU threshold of 0.7, which exceeds many existing conventional deep learning models, while the lightweight design of our proposed module is more helpful for the implementation of artificial intelligence technology in the field of underwater wireless communication.