改进的DeepLabCut鱼类游动轨迹提取
Improved DeepLabCut for fish trajectory extraction
雷帮军 1裴斐 1吴正平 1张海镔1
作者信息
- 1. 湖北省水电工程智能视觉监测重点实验室,湖北 宜昌,443002;三峡大学计算机与信息学院,湖北 宜昌,443002;水电工程视觉监测宜昌市重点实验室,湖北 宜昌,443002
- 折叠
摘要
针对现有的鱼类游动轨迹提取方法在提取效率和准确率方面不能同时兼顾的问题,提出了一种改进的DeepLabCut方法用于鱼类背部关键点识别和定位.首先,选择了轻量级卷积神经网络模型EfficientNet-B0作为DeepLabCut的主干网络模型,用于提取鱼类背部关键点的特征,为了增强EfficientNet-B0 的表征能力,在网络模型中引入了改进的CBAM(Convolutional Block Attention Module)注意力机制模块,将CBAM中的空间注意力模块和通道注意力模块从原来的串行连接方式改为并行连接,以解决两种注意力模块之间因串行连接而导致的互相干扰问题.其次,基于MSE(Mean Squared Error)损失函数提出了一种分段式损失函数H_MSE用于模型的训练,分段式损失函数H_MSE相对于传统的损失函数具有较强的鲁棒性,其在处理数据中的异常值时能表现出较低的敏感性.最后,采用了半监督学习方法对关键点进行自动标记来减少人工标记数据时产生的误差.结果显示:相比于DeepLabCut原始算法,识别误差RMSE(Root Mean Squared Error)平均降低了 4.5 像素;与目标检测算法Faster RCNN、SK-YOLOv5、ESB-YOLO、YOLOv8-Head-ECAM相比,识别误差RMSE平均降低了 11.5 像素,检测效果优于其他目标检测网络和原始网络,平均每张图像的检测时间为 0.062 s,能够快速准确提取鱼道内鱼类的游动轨迹,为优化鱼道的水力设计指标提供了重要依据.
Abstract
A modified DeepLabCut method is proposed to address the problem of existing fish trajectory extraction methods that cannot simultaneously balance efficiency and accuracy.Firstly,the lightweight convolutional neural network model EfficientNet-B0 is chosen as the backbone model of DeepLabCut for extracting key points on the back of fish.To enhance the representation capability of EfficientNet-B0,an improved Convolutional Block Attention Module(CBAM)is introduced into the network model.The spatial and channel attention modules in CBAM are connected in parallel instead of the original sequential connection to solve the mutual interference problem caused by the sequential connection between the two attention modules.Secondly,a segmented loss function H_MSE based on Mean Squared Error(MSE)is proposed for model training.The segmented loss function H_MSE exhibits strong robustness compared to traditional loss functions,showing lower sensitivity when handling outliers in the data.Finally,a semi-supervised learning approach is adopted to automatically label the key points,reducing errors caused by manual labeling.The results show that compared to the original DeepLabCut algorithm,the average root mean squared error(RMSE)of recognition is reduced by 4.5 pixels.Compared to object detection algorithms such as Faster RCNN,SK-YOLOv5,ESB-YOLO,and YOLOv8-Head-ECAM,the average RMSE recognition error is reduced by 11.5 pixels.The detection performance is superior to other object detection networks and the original network,with an average detection time of 0.062 s per image.It can quickly and accurately extract the swimming trajectories of fish in the fish passage,providing an important basis for optimizing the hydraulic design indicators of the fish passage.
关键词
鱼类识别/轨迹识别/关键点识别/DeepLabCut/半监督学习/损失函数/注意力机制Key words
fish recognition/trajectory recognition/keypoint recognition/DeepLabCut/semi-supervised learning/loss function/attention mechanism引用本文复制引用
基金项目
国家自然科学基金(61871258)
水电工程智能视觉监测湖北省重点实验室建设项目(2019ZYYD007)
出版年
2024
国家科技期刊平台
NSTL
万方数据