SENet优化的Deeplabv3+淡水鱼体语义分割

淡水鱼头、腹、鳍的各部分快速识别与精准定位是机器人实现淡水鱼快速抓取,精确切割、提升作业效率关键技术的前提。针对深度学习的淡水鱼体语义分割算法在编码特征提取阶段产生大量无效的特征通道,以及网络不断下采样和池化操作使得鱼体某些细节信息被丢失,网络性能下降、边缘分割效果不佳的问题,提出了一种基于SENet优化后的Deeplabv3+淡水鱼头、腹、鳍的语义分割算法。利用空洞/带孔卷积(dilated/atrous convolutions)实现扩展感受野,克服细节信息丢失,达到准确定位的目的,同时SENet的优化使得Deeplabv3+通过学习的方式提升淡水鱼有用的特征并抑制对当前任务用处不大的特征,最终淡水鱼各部分的语义分割平均交并比(MIoU)在自建的淡水鱼数据集上达到了93%左右,性能得到了明显提升并达到了目前先进分割水平。

SENet optimized Deeplabv3+freshwater fish body semantic segmentation

The rapid identification and precise positioning of the freshwater fish head,belly,and fins are the prerequisites for the robot to realize the rapid grasping of freshwater fish,precise cutting,and the key technology of improving operation efficiency.The freshwater fish body semantic segmentation algorithm for deep learning produces a large number of invalid feature channels in the encoding feature extraction stage,and the continuous down-sampling and pooling operations of the network make certain details of the fish body lost,the network performance is reduced,and the edge segmentation effect is not good of the problems,a semantic segmentation algorithm based on Deeplabv3+freshwater fish head,belly and fin optimized due to SENet is proposed.Use dilated/atrous convolutions to expand the receptive field,overcome the loss of detailed information,and achieve accurate positioning.At the same time,the optimization of SENet enables Deeplabv3+to improve the useful features of freshwater fish through learning and suppress the current task of features that are not very useful,the final semantic segmentation mean intersection ratio(MIoU)of each part of freshwater fish reached about 93%on the self-built freshwater fish data set,and the performance was significantly improved and reached the current advanced segmentation level.