基于BMM⁃Fast⁃SCNN的海底线实时提取算法  

Real⁃Time Extraction Algorithm of Bottom Based on BMM⁃Fast⁃SCNN

作　　者：陈朋[1] 昝昊 赵冬冬 郭新新 CHEN Peng;ZAN Hao;ZHAO Dongdong;GUO Xinxin(College of Computer Science&Technology,Zhejiang University of Technology,Hangzhou Zhejiang 310023,China;College of Information Engineering,Zhejiang University of Technology,Hangzhou Zhejiang 310023,China;Institute of Deep-sea Science and Engineering,Chinese Academy of Science,Sanya Hainan 572000,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,浙江杭州310000 [2]浙江工业大学信息工程学院,浙江杭州310000 [3]中国科学院深海科学与工程研究所,海南三亚572000

出　　处：《传感技术学报》2023年第8期1227-1234,共8页Chinese Journal of Sensors and Actuators

基　　金：国家自然科学基金青年科学基金项目(62001418);浙江省自然科学基金项目(LQ21F010011);中国科学院战略性先导科技专项项目(A类)(XDA22030302)。

摘　　要：侧扫声呐瀑布图由于其成像原理,需要借助海底线位置对原始图片进行斜距校正。然而在实际采集过程中,声呐自噪声、悬浮物和海底目标等许多干扰因素会增加海底线提取的难度,现有的传统方法和神经网络方法在回波信号信噪比较差时无法正确实时地提取海底线。针对这些问题,提出了一种基于海底信息对称模块和多尺度特征融合模块的快速分割卷积神经网络(Bottom Information Symmetry Module and Multi⁃scale Feature Fusion Module Fast⁃SCNN,BMM⁃Fast⁃SCNN)用于实时正确提取海底线。该算法基于Fast⁃SCNN⁃1D,结合海底信息对称模块(Bottom Information Symmetry Module,BISM)来提高网络的鲁棒性,并通过多尺度特征融合模块(Multi⁃scale Feature Fusion Module,MFFM)增强网络提取海底线细小特征的能力。在两条测线中,该算法在一个像素和两个像素误差范围内的提取精度分别为83.56%,97.63%和96.27%,99.49%,相较于其他方法,分别至少提高了1.41%、9.07%和0.74%、0.48%。实验结果表明,该算法能够有效地提取海底线,方法执行时间为9.38 ms,满足实时正确提取海底线的要求。Due to its imaging principle,the side⁃scan sonar waterfall needs to use the position of the bottom to correct the slant distance of the original image.However,in the actual acquisition process,many interference factors such as sonar self⁃noise,suspended objects and seabed targets will increase the difficulty of bottom extraction.The existing traditional methods and neural network methods cannot correct⁃ly extract the bottom in real time when the signal⁃to⁃noise ratio of the echo signal is poor.Aiming at these problems.A fast segmentation convolutional neural network based on bottom information symmetry module and the multi⁃scale feature fusion module(BMM⁃Fast⁃SCNN)is proposed to correctly extract bottom in real time.The algorithm is based on Fast⁃SCNN⁃1D,combined with bottom information symmetry module(BISM)to improve the robustness of the network,and enhanced the ability of the network to extract small features of the bottom through multi⁃scale feature fusion module(MFFM).In the two survey lines,the values of extraction accuracy of the algorithm within the error values range of one pixel and two pixels are 83.56%,97.63%and 96.27%,99.49%,which are at least 1.41%,9.07%and 0.74%,0.48%higher than those of other methods.The experimental results show that the algorithm can effectively extract the bottom,and the exe⁃cution time of the method is 9.38 ms,which meets the requirements of real⁃time and correct extraction of the bottom.

关 键 词：图像处理 海底线提取 卷积神经网络 BISM 声呐图像 

分 类 号：P229.1[天文地球—大地测量学与测量工程]