基于形状及纹理特征的淡水鱼种类自动识别方法  

An automatic method for freshwater fish species classification using shape and texture features

作　　者：梁钊董 熊兴国 徐东坡[2] 陆明洲[1] 沈明霞[1] 张婉平 童奇烈 LIANG Zhaodong;XIONG Xingguo;XU Dongpo;LU Mingzhou;SHEN Mingxia;ZHANG Wanping;TONG Qilie(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210031,China;Freshwater Fisheries Research Center,Chinese Academy of Fishery Sciences,Wuxi 214128,China;Supervision and Administration Station of Hangzhou for Fishing Industry,Vessels,and Ports,Hangzhou 310008,China)

机构地区：[1]南京农业大学人工智能学院,江苏南京210031 [2]中国水产科学研究院淡水渔业研究中心,江苏无锡214128 [3]杭州市渔政渔港渔船监督管理总站,浙江杭州310008

出　　处：《南京农业大学学报》2021年第3期576-585,共10页Journal of Nanjing Agricultural University

基　　金：浙江省杭州市渔政总站萧山湘湖水生生物资源调查项目(CTZB-F190435IWZ-Z)。

摘　　要：[目的]本文利用自行搭建的淡水鱼图像采集装置在淡水鱼捕捞现场采集鱼体图像,在图像数据集上提出一种实用的淡水鱼自动识别分类方法。[方法]对14类淡水鱼图像进行预处理、自模板匹配操作,分割出完整的鱼体前景图像;提取鱼体全身、去尾鱼体及鱼尾的16维形状特征,利用灰度-梯度共生矩阵(GLGCM)、局部二值模式(LBP)及Gabor变换提取鱼体纹理特征,采用主成分分析法(PCA)分别筛选GLGCM纹理+形状、LBP纹理+形状和Gabor纹理+形状特征累积贡献率超过85%的特征组合。将降维前后特征集的70%和30%分别作为训练集和验证集,利用朴素贝叶斯、K近邻(KNN)、线性回归、决策树、随机森林、支持向量机(SVM)、梯度提升决策树(GBDT)7种机器学习方法训练淡水鱼品种分类器,并利用验证集数据分析对比各分类器的性能。[结果]鱼体前景图像分割算法测试结果表明,本文提出的自模板匹配方法可在不建立大规模模板库的前提下,以99.79%的正确率分割鱼体图像。分类器性能验证及对比结果表明,基于GLGCM纹理+形状特征的随机森林分类器的淡水鱼识别精度最高,降维获得的5维GLGCM纹理+形状特征向量识别14种淡水鱼的正确率达到99.52%。[结论]提出的自模板匹配方法可以在不构建庞大模板库的前提下实现鱼体前景区域的分割,基于筛选得到的GLGCM纹理+形状特征的随机森林分类器可用于自动识别淡水鱼品种。[Objectives]An automatic method for freshwater fish species classification was proposed in this paper,based on the fish images collected by using a homemade fish image acquisition device.[Methods]A self-template matching method was developed to segment the fish body region from the binary image,which was obtained after the image pre-processing operations.Then the shape and texture features of the fish body were extracted,where the latter were obtained using three methods,including gray level-gradient co-occurrence matrix(GLGCM),local binary pattern(LBP),and Gabor filter.Principal component analysis(PCA)was carried out to transform a high dimensional feature set of GLGCM texture and shape,LBP texture and shape,or Gabor texture and shape into smaller ones that contained at least 85%of the information in the large feature set.70%and 30%of the feature sets were used to train and validate species classification models established using seven machine learning methods[Naive Bayes,K-neares neighbors(KNN),linear regression,randow forest,decision tree,support vector machine(SVM),gradient boosting decision tree(GBDT)].[Results]Fish body segmentation test indicated that the developed self-template matching method could segment the region on interest with a correct rate of 99.79%,which was achieved without the need of an extra fish body template library.Classification models validation results indicated that the random forest classifier using reduced dimensional GLGCM texture and shape feature performed best,and an average accuracy of 99.52%was achieved when it was used to classify all the 14 freshwater fish species.[Conclusions]The self-template matching method developed in this study could extract fish body region well without the need of an extra fish body template library.The reduced dimensional GLGCM texture and shape feature set was suitable for identifying the species of freshwater fish.

关 键 词：机器视觉 淡水鱼分类 纹理 特征降维 随机森林 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]