摘要
水下环境鱼类分割是实现体长测量、体重估算和群体计数等智能化测量的关键技术,为了提高分割的准确性,提出一种融合边缘监督的改进Deeplabv3+鱼类分割方法。编码部分采用更少的下采样次数,浅层增加卷积块注意力机制(convolutional block attention module,CBAM),以减少信息丢失并增强浅层语义信息;通过设计混合膨胀卷积(hybrid dilated convolution,HDC)改进空洞空间卷积池化金字塔(atrous spatial pyramid pooling,ASPP)模块,提取深层特征;在解码输出部分结合Canny边缘检测算子引入边缘监督,通过边缘损失函数来获得边缘预测和边缘标签的误差以更好地学习边缘特征;最后根据不同类像素比率引入优化的损失函数,进一步提升模型的语义分割性能。该方法在VOC2012数据集上mIoU达到84.56%,较Deeplabv3+方法提升了3.27%,验证了其泛化能力。在DeepFish数据集上做消融实验,mIoU高达93.66%,均高于Deeplabv3+、Unet和PSPNet等常见方法。该研究提升了水下环境鱼类分割的精度,可为水产养殖智能化提供支持。
Fish segmentation in underwater environment is the key technology to realize intelligent measurement such as body length measurement,weight estimation and population counting.In order to improve the accuracy of fish segmentation,an improved Deeplabv3+fish segmentation method combined with edge supervision is proposed.In the encoder part,fewer down sampling times are used,and convolutional block attention module(CBAM)is added in the shallow layer to reduce information loss and enhance the shallow semantic information;By designing hybrid dilated convolution(HDC)to improve atrous spatial pyramid pooling(ASPP)module,deep features are extracted.In the decoder output part,Canny edge detection operator is combined to introduce edge supervision,and the edge prediction and edge label errors are obtained through the edge loss function to better learn edge features.Finally,the optimized loss function is introduced according to different pixel ratios to further improve the semantic segmentation performance of the model.This method achieves 84.56%mIoU on VOC2012 dataset,which is 3.27%higher than Deeplabv3+method,and verifies its generalization ability.In the ablation experiment on DeepFish dataset,mIoU is as high as 93.66%,which is higher than common methods such as Deeplabv3+,Unet and PSPNet.This research improves the accuracy of fish segmentation in underwater environment and can provide support for intelligent aquaculture.
作者
田志新
廖薇
茅健
吴建民
袁泉
徐震
Tian Zhixin;Liao Wei;Mao Jian;Wu Jianmin;Yuan Quan;Xu Zhen(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;Shanghai Academy of Agricultural Sciences,Shanghai 201400,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2022年第10期208-216,共9页
Journal of Electronic Measurement and Instrumentation
基金
国家农业环境奉贤观测实验站项目(NAES035AE03)
上海市科技兴农项目(2022-02-08-00-12-F01186)
国家自然科学基金青年基金项目(62001282)资助
作者简介
田志新,2018年于河北工业大学获得学士学位,现为上海工程技术大学硕士生,主要研究方向为图像处理、计算机视觉及其应用。E-mail:2212720600@qq.com;通信作者:徐震,2006年于华东师范大学获得学士学位,2013年获得上海大学博士学位,2014年1月到2016年1月在中科院上海应用物理研究所从事博士后研究,现为上海工程技术大学讲师,主要研究方向为数据驱动的机器学习和计算机视觉算法发展及其应用。E-mail:lcxuzhen@163.com
