基于多特征的淡水鱼种类识别研究  

Species recognition of fishes based on multiple features

作　　者：谢忠红[1] 郭小清[1] 程碧云 李姣[1] 朱淑鑫[1] 徐焕良[1] 

机构地区：[1]南京农业大学信息科学技术学院,江苏南京210095

出　　处：《扬州大学学报（农业与生命科学版）》2016年第3期71-77,共7页Journal of Yangzhou University：Agricultural and Life Science Edition

基　　金：中央高校基本科研业务费项目(KYZ201670;KYZ201551);江苏省科技支撑计划项目(BE2011339)

摘　　要：为促进淡水鱼产业的发展、提高淡水鱼机器视觉的识别效率,以4种常见的淡水鱼(鲫鱼、桂鱼、鳊鱼和白鱼)为研究对象,提出一种基于形态、纹理和颜色等多特征融合的淡水鱼种类识别方法。将淡水鱼图像转换至HIS空间后,对S分量图像利用ostu分割方法获取鱼体二值图像,计算鱼体的多个形态特征(面积/周长、复杂度、偏心率、长宽比和7个Hu不变矩)。针对4种鱼尾巴形态差异较大的特点,引入尾巴形态特征进行种类识别,按列扫描鱼体二值图像获取鱼体宽度函数,先对宽度函数进行傅里叶变换,找到鱼体最小宽度位置后分割出鱼尾,计算鱼尾形态特征(占空比、似圆度)。根据各种淡水鱼的鱼肚、鱼背皮肤纹理和颜色存在较大差异的客观性,提出基于鱼背和鱼肚2块皮肤的特征抽取方法,抽取22个特征;加上鱼体和鱼尾的13个形态特征,总计35个特征,选择使用粒子群算法优选出22个特征后利用改进的遗传神经网络法对淡水鱼进行种类识别。识别样本集R中共240条鱼2组图像,第1组开启辅助光源,第2组采用自然光照。设置一条鱼纹理块识别正确率的阈值T为80%时,第1组60条鳊鱼和60条桂鱼全部识别成功,识别正确率为100%,60条鲫鱼和60条白鱼各识别出59条,识别正确率为98.3%。第2组图像在2种天气情况下拍摄,晴天拍摄140条鱼(每种35条),鳊鱼、桂鱼和白鱼的识别正确率为100%,而鲫鱼识别出34条,识别正确率为97.1%;阴天拍摄100条鱼(每种25条),纹理子块的识别正确率均大于T的条数为23条,识别正确率23%,说明阴雨天无辅助光源时由于光线过于昏暗,特征值不明显,导致整体识别正确率很低。这一研究提出的多特征融合的淡水鱼种类识别方法适用于有辅助光源或晴天光线好的环境,但光线暗的阴雨天不适用,这为后期研制淡水鱼在线识别装置提供了理论依据。In order to promote the development of freshwater fish industry,a species recognition method for 4types of freshwater fishes based on multiple features including shape,texture and color,etc was proposed.At first the fish image was conversed into HIS color space,and S component image in HIS color space was processed by enhancing,denoising,corroding,expanding,etc.The double value image of fish was segmented from S component image by use of OSTU segmentation method.At last the fish shape characters（complexity,eccentricity,aspect ratio and 7 Hu＇s invariants）were calculated.Four fish tail shapes were found out different by close observation,so tail＇s shape characters were considered very important for species recognition.After the fish tail was cut from fish body at the position of minimum body width which was found by using Fourier transform in fish＇s width function,the shape characters of fish tail（duty ratio and roundness）were calculated.Comparing skins of fish tummy and fish back,the difference was very obvious.So this study creatively proposed a character extraction method which was based on characters of two areas of fish skin.After two areas of skin（300pixel×150pixel）cut from fish tummy and back were transferred to 3types of space：Gray space,HSV space and HIS space,textural features such as energy,contrast,correlation and entropy were extracted from image imgGray and image imgI1 based on GLCM（Gray level co-occurrence matrix）.H,Sand V color components were also extracted from HSV color space.Added by the 13 characters of fish tummy and fish back.22 characters which were selected based on particle swarm optimization algorithm through comparing with genetic algorithm were input into improved genetic neural network to recognize the fish species.The fish tummy and fish back were divided into 72 little blocks whose size was 32pixel×32pixel,and 22 characters including color,texture and shape based on every piece of skin and shapes of fish body and tail were input into improved genetic neural

关 键 词：淡水鱼 机器视觉 特征选择 种类识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]