构建VED-SegNet分割模型提取鱼类表型比例

为提高鱼类表型分割精度和准确度,实现鱼类表型智能监测,该研究基于深度学习算法构建了VED-SegNet模型用于鱼类表型分割和测量。该模型将cross stage partial network和GSConv结合作为编码器(VoV-GSCSP),保持足够精度的同时降低网络结构复杂性。另一方面,该模型采用EMA(efficient multi-scale attention module with cross-spatial learning)建立强化结构,加强编码器和解码器之间的信息传递,提高模型精度,并实现了8个表型类别的输出。采用自建的鱼类表型分割数据集对VED-SegNet模型进行了测试,测量结果中鱼类各表型比例与实际测量值相接近,表型最大平均绝对和平均相对误差为0.39%、11.28%,能实现无接触式提取水产养殖中鱼类表型比例。对比其他常见语义分割模型,平均交并比mean intersection over union,mIoU和平均像素准确率mean pixel accuracy,m PA最高,分别到达了87.92%、92.83%。VED-SegNet模型在环境复杂、多鱼重叠的养殖水体鱼类监测场景中精准分割鱼类形态特征,可为鱼类表型智能测量提供技术支持。

Relationship between trawl selectivity and fish body size in a simulated population

Trawl is a main fishing gear in Chinese fishery,capturing large fish and letting small ones at large.However,long-term use of trawl would result in changes of phenotypic traits of the fish stocks,such as smaller size-at-age and earlier age-at-maturation.In this study,we simulated a fish population with size characteristics of trawl fishing and the population produces one generation of offspring and lives for one year,used trawl to exploit the simulated fish population,and captured individuals by body size.We evaluated the impact of the changes on selectivity parameters,such as selective range and the length at 50% retention.Under fishing pressure,we specified the selectivity parameters,and determined that smaller selection rates and greater length at 50% retention were associated with an increased tendency towards miniaturization.