海洋渔业2026，Vol.48Issue(4)：540-550,11.
基于改进DeepLabV3+和迁移学习的鱼体表型信息提取方法
Fish phenotype information extraction method based on improved DeepLabV3+and transfer learning
摘要
Abstract
Semantic segmentation of fish body images is a key technology for detecting fish body surface characteristic parameters and group behavior.In order to improve segmentation accuracy,an improved DeepLabV3+fish body image segmentation model was proposed.The model integrated the lightweight module(MobileNetV3),introduced the attention module(ECA)and improved the the atrous spatial pyramid pooling(ASPP)structure and adopted transfer learning training methods.The training results of the classical U-Net,SegNet,PSPNet and the improved DeepLabV3+semantic segmentation network in the fish body image data set were compared.Experimental results showed that the improved DeepLabV3+network model had higher segmentation accuracy,feature extraction and inference efficiency in the integrated fish body image data set,with the mean intersection over union(MIoU)reaching 94.61%and the accuracy reaching 93.78%.A fish body model was used to detect the parameters of total length,body height,perimeter and area,and the error range between the average value and the true value was less than 3.74%.The improved DeepLabV3+network model elevated the accuracy of semantic segmentation of fish body image,and could provide a reference for the automatic measurement of fish body phenotype characteristic parameters.关键词
机器视觉/深度学习/图像分割/DeepLabV3+/鱼体表型Key words
machine vision/deep learning/image segmentation/DeepLabV3+/fish phenotype分类
信息技术与安全科学引用本文复制引用
王军溢,张俊,曹守启,周普洛,徐畅..基于改进DeepLabV3+和迁移学习的鱼体表型信息提取方法[J].海洋渔业,2026,48(4):540-550,11.基金项目
国家重点研发计划项目(2023YFD2400502) （2023YFD2400502）
国家自然科学基金(32573573) （32573573）
上海市自然科学基金(24ZR1429000) （24ZR1429000）
