基于Faster R-CNN的对虾生长性状表型高通量测定技术的建立及应用  

THE ESTABLISHMENT AND APPLICATION OF A FAST PHENOTYPIC DETERMINATION TECHNIQUE BASED ON FASTER R-CNN FOR GROWTH TRAITS IN SHRIMP

作　　者：鲍镇宁 于洋[1] 李富花[1] BAO Zhen-Ning;YU Yang;LI FU-Hua(Key Laboratory of Experimental Marine Biology,Institute of Oceanology,Chinese Academy of Sciences,Qingdao 266071,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院海洋研究所实验海洋生物学重点实验室,青岛266071 [2]中国科学院大学,北京100049

出　　处：《水生生物学报》2023年第10期1576-1584,共9页Acta Hydrobiologica Sinica

基　　金：中国科学院战略性先导科技专项(XDA24030105)资助。

摘　　要：为提高对虾外部生长性状表型数据的获取效率,利用拍照获得的对虾外部表型照片,采用基于区域生成网络RPN(Region Proposal Networks)的Faster R-CNN(Faster Region-convolutional neural networks)深度学习神经网络,通过对8400张对虾表型照片的学习和训练,构建了快速识别对虾全长并输出位置信息的模型。该模型可识别图片中的对虾并以识别框的形式表示出具体的位置。对于不同角度拍摄的对虾,模型生成识别框的长度或对角线长度与人工测量的对虾全长之间呈高度相关。研究以此建立了对虾全长性状表型数据高通量测定技术,该技术的建立可以在对虾生长性状表型数据测定中节省人工测量的时间,提高了对虾全基因组选择育种的效率。此外,该模型的建立也为对虾头胸部长度及不同体节长度等其他外部表型数据的测定提供了新的思路,为对虾生长性状表型组的建立奠定了重要基础。High-throughput genotyping and phenotyping are the key techniques for efficient and precise breeding.Compared with genotyping technology,the development of high-throughput phenotyping is relatively slower.In breeding of aquatic animals,especially for shrimp,the phenotypic data of the external growth traits are mainly obtained by manual measurement,which needs high labor intensity with low efficiency.In recent years,the rapid development of deep learning has provided technical support for high-throughput phenotyping.In order to improve the efficiency of obtaining phenotypic data of external growth traits of shrimp,we applied a Faster R-CNN(Faster Region-convolutional neural networks)model based on Region Proposal Networks(RPN)to automatically identify the body length of shrimp.Based on the training to 8400 shrimp photos,the model could rapidly identify and output location information of shrimp,and the total lengths of shrimp were accurately measured.For the shrimp photos taken in vertical view,the length of the recognition frame was highly correlated with the manually measured full length of the shrimp.For the photos taken in side view,some individuals of shrimp were in bent shape which will affect the correlations.We found that the ratio of the diagonal length of the recognition frame to the length(K value)could represent the degree of bending of the shrimp.Further analysis illustrated that the shrimp is straight in the picture and the length of the recognition frame is highly correlated with the full length of the shrimp when the K is less than 1.04.However,most of the shrimp in the photo is bent,and the diagonal length of the recognition frame is highly correlated with the full length of the shrimp,when the K is greater than 1.04.Consequently,we established a high-throughput technique to determine the full-length of shrimp.The establishment of this technique can save time in comparison to manual measurements on the shrimp phenotype and improve the efficiency of genomic selection breeding of shrimp.In addition,t

关 键 词：表型组 计算机视觉 深度学习 对虾 

分 类 号：Q-332[生物学]