基于语义部位分割的条纹斑竹鲨鱼体运动姿态解析  

Motion posture parsing of Chiloscyllium plagiosum fish body based on semantic part segmentation

作　　者：刘斌[1] 王凯歌 李晓蒙 胡春海[1] Liu Bin;Wang Kaige;Li Xiaomeng;Hu Chunhai(Hebei Key Laboratory for Test and Measurement Technology,Yanshan University,Qinhuangdao 066004,China)

机构地区：[1]燕山大学河北省测试测量技术重点实验室,秦皇岛066004

出　　处：《农业工程学报》2021年第3期179-187,共9页Transactions of the Chinese Society of Agricultural Engineering

基　　金：河北省高等学校科学技术研究项目(15220324)。

摘　　要：条纹斑竹鲨具有较高的经济价值和医用研究价值。人工驯养对环境和温度等因素要求较高,时常出现大规模病死现象。利用视频图像量化分析鱼体运动行为,有助于进行异常识别和早期预警,将有效提高养殖养护水平。该研究针对人工驯养的条纹斑竹鲨鱼,提出一种基于深度神经网络的语义部位分割方法,并将分割结果应用于剖析条斑鲨鱼体运动姿态。首先,依据条斑鲨形态特征将其划分为7个可视的身体组成构件(头部、右胸鳍、左胸鳍、右腹鳍、左腹鳍、躯干、尾巴);再对全景养殖监控视频中抽取的476幅条斑鲨子图进行各部位的像素级标记,通过数据增强到1944幅建立鱼体语义部位数据集,其中训练集为1166幅图像,测试集为778幅图像;然后,在语义分割网络模型基础上进行深度学习训练,使用深度学习框架对网络参数进行微调使得网络训练结果达到最优。最后,利用语义部位分割结果定位躯干和鱼头质心建立随体坐标,通过随体坐标的方向变化判明鱼体动作姿态。基于FCN-8s和Segnet两种深度网络模型进行了鱼体部位分割的对比试验,测试结果表明基于Segnet网络的分割方法在头部、右胸鳍、左胸鳍、右腹鳍、左腹鳍、躯干、尾巴部位的准确度分别高出FCN-8s深度网络1.50,4.70,6.95,6.56,6.01,0.85,0.84个百分点。语义部位分割结果能够有效判别条斑鲨鱼体目标的动作姿态,可为鱼体异常行为识别和进一步开展面向条斑鲨的动物行为学试验提供技术参考。The Chiloscyllium plagiosum has high economic and medical value.However,the real artificial breeding conditions cannot meet the high requirements for the breeding environment of marine fish,such as water quality and temperature,often leading to large-scale illness even death.Since video imaging has been widely used to quantitatively analyze the movement behavior of farmed fish,the technique can contribute to identifying abnormal behavior for the early warning,and thereby effectively improving the level of breeding and conservation.In this study,an imaging algorithm was proposed for the semantic part segmentation of Chiloscyllium plagiosum using encoder-decoder architecture,thereby analyzing the body movement and posture of the Chiloscyllium plagiosum.Three steps were as follows:1)The images of Chiloscyllium plagiosum were divided into 7 visible body components,according to the morphological characteristics,including the head,left pectoral fin,right pectoral fin,left ventral fin,right ventral fin,trunk,and tail.Then,the sub-images of Chiloscyllium plagiosum were extracted from the video images in the panoramic breeding surveillance under a breeding circumstance,where a total of 476 candidate patterns were obtained,while all the images in the dataset were manually marked.After that,data augmentation was used to increase the number of images,and thus a total of 1944 images were obtained,of which 1166 images were selected as training images,and 778 images were selected as test images.2)The pre-processed training dataset was fed into the network model of semantic segmentation by fine-tuning network parameters,where a deep learning framework was used to optimize the network training for the best.Then,the test dataset was put into the trained model for the segmentation.3)Post-processing was performed to fill the holes within objects or remove small objects,where a disk structure of mathematical morphology was used to calculate the areas of connected regions.Simple and effective post-processing was utilized to obtain the

关 键 词：机器视觉 神经网络 条纹斑竹鲨 语义部位分割 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]