船舶目标重叠下马赛克图像数据增强方法研究  

Research on mosaic image data enhancement and detection method for overlapping ship targets

作　　者：曾广淼 俞万能[1,2] 王荣杰 林安辉[1,2] ZENG Guang-miao;YU Wan-neng;WANG Rong-jie;LIN An-hui(School of Marine Engineering,Jimei University,Xiamen Fujian 361021,China;Fujian Provincial Key Laboratory of Naval Architecture and Ocean Engineering,Xiamen Fujian 361021,China)

机构地区：[1]集美大学轮机工程学院,福建厦门361021 [2]福建省船舶与海洋工程重点实验室,福建厦门361021

出　　处：《控制理论与应用》2022年第6期1139-1148,共10页Control Theory & Applications

基　　金：国家自然科学基金项目(51879118,52171308);福建省自然科学基金项目(2020J01688);福建省科技重点项目(2021H0021);福建省科技拥军项目(B19101);交通运输行业高层次技术人才培养项目(2019–014);集美大学青年拔尖人才项目(ZR2019006)资助。

摘　　要：目标识别中的重叠遮挡问题一直以来是研究的难点,船舶目标在狭窄水域发生相互遮挡的情况依然存在.本文提出了一种改进的马赛克数据增强方法,将训练图片拼接变换成3种不同的尺度,并按照不同比例输入网络进行训练,强化了检测算法对局部特征的学习能力,在保持测试速度不变的情况下,提高了对重叠目标的识别准确率,降低了不同分辨率下识别能力的衰弱速度,加强了算法的鲁棒性.基于小型移动测试平台的实测实验证明,相对于原始算法,经过改进后的算法在重叠目标的识别准确率上提高了2.5%,目标丢失时间减少了17%,在不同视频分辨率下的识别稳定性上提高了27.01%.The problem of overlapping occlusion in target recognition has been a difficult research problem, and the situation of mutual occlusion of ship targets in narrow waters still exists. In this paper, we proposed an improved mosaic data enhancement method, which transforms the training image stitching into three different scales and inputs them into the network for training at different scales to strengthen the learning ability of the detection algorithm for local features.While keeping the test speed constant, the recognition accuracy of overlapping targets is improved, the rate of decay of recognition ability at different resolutions is reduced, and the robustness of the algorithm is enhanced. The experiments based on a small mobile testbed proved that, compared to the original algorithm, the improved algorithm improved the recognition accuracy of overlapping targets by 2.5%, reduced the target loss time by 17%, and improved the recognition stability at different video resolutions by 27.01%.

关 键 词：船舶识别 目标重叠 图像数据增强 Yolov4算法 深度学习 

分 类 号：U675.7[交通运输工程—船舶及航道工程] TP391.41[交通运输工程—船舶与海洋工程]