深度语义分割MRF模型的海洋筏式养殖信息提取  

Deep semantic segmentation MRF model for information extraction of marine floating raft aquaculture

作　　者：王心哲 邓棋文 王际潮 范剑超[3] WANG Xinzhe;DENG Qiwen;WANG Jichao;FAN Jianchao(School of Information Science and Engineering,Dalian Polytechnic University,Dalian 116034,Liaoning,China;School of Data Science,City University of Hong Kong,Hong Kong 999077,China;Department of Marine Remote Sensing Technology,National Marine Environment Monitoring Center,Dalian 116023,Liaoning,China)

机构地区：[1]大连工业大学信息科学与工程学院,辽宁大连116034 [2]香港城市大学数据科学学院,中国香港999077 [3]国家海洋环境监测中心海洋遥感技术室,辽宁大连116023

出　　处：《山东大学学报（工学版）》2022年第2期89-98,共10页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金项目(42076184,41876109,41706195);高分重大科研专项(41-Y30F07-9001-20/22);国家重点研发计划(2017YFC1404902,2016YFC1401007)。

摘　　要：采用无监督方法与深度学习模型结合,解决筏式养殖边缘信息精确提取问题,提出深度语义分割(semantic segmentation, SegNet)-马尔科夫随机场(Markov random field, MRF)模型,该模型提取目标空间细节信息和深度判别特征信息。通过SegNet编码器的卷积和最大池化提取筏式养殖的特征信息和扩大感受野,抑制噪声、误判等现象的产生,模型后端接入MRF模型,计算像素空间领域内的特征信息进行聚类分析来获取目标低水平的空间细节信息,在深度特征信息的基础上较大程度的保留空间特征信息,完善边缘信息并抑制连通区域的产生。试验结果表明,该模型极大减少了特征信息丢失和因海水背景而产生的误判,其分类精度高于95%,明显优于经典无监督算法和单一的深度学习模型。The unsupervised method combined with deep learning model was adopted to solve the problem of accurate extraction of raft aquaculture edge information, a method of information extraction combining deep learning SegNet(semantic segmentation) and MRF(Markov random field) was proposed. In this method, two models were used to obtain the spatial detail information and depth discrimination feature information. The convolution and maximized pooling of SegNet encoder were used to extract the feature information of raft aquaculture and expand receptive fields to suppress noise and misjudgment. MRF was connected to the back end of the model, the feature information in the pixel space domain was calculated for clustering analysis to obtain the low-level spatial details of the target. Based on the depth feature information, the spatial feature information was largely retained, the edge information was improved, and the generation of connected regions was suppressed. The experimental results showed that the model could greatly reduce the loss of feature information and misjudgment caused by the sea background, the proposed classification accuracy was higher than 95%, which was obviously better than the classical unsupervised algorithm and single deep learning model.

关 键 词：筏式养殖 卷积神经网络 深度学习 马尔科夫随机场 遥感影像 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]