基于语义分割算法的水库漂浮物智能识别

Intelligent recognition of floating debris in reservoirs based on semantic segmentation algorithm

为了满足水库漂浮物强监管的需求,发挥水库水雨情视频监控的优势,将语义分割算法DeepLabV3+应用于水库漂浮物智能识别任务中。首先,通过多要素数据采集方法以及多种图像数据增强方式系统性整合得到水库漂浮物语义分割数据集。然后,提出改进的加权代价函数,以适应数据集中各类的总像素数量不平衡的问题。实验结果表明,通过规范采集与多种数据增强建立的数据集,以及改进的加权代价函数,提高了水库漂浮物语义分割模型的准确率。所提模型像素准确率平均值、平均交叉比分别可达到95.21%、91.03%,通过微信企业号实现的移动监管,可为水利相关部门对水库漂浮物进行及时的智能监控预警提供新的技术方案。

To satisfy the demand for strong supervision of reservoir floating debris and get the utmost out of the advantages of video monitoring water and rain conditions in the reservoirs,the semantic segmentation algorithm DeepLabV3+was applied to the intelligent recognition of floating debris in reservoirs.Firstly,the semantic segmentation data set of floating debris in the reservoirs was systematically integrated through the multi-element data collection method and multiple image data augmentation methods.Then,an improved weighted cost function was proposed to solve the imbalance problem of the total pixel number among different classes in the data set.The experimental results show that,the accuracy of the semantic segmentation model of floating debris in reservoirs is improved by the data set established by standardized collection and multiple data augmentation,and the improved weighted cost function.The mean pixel accuracy and the mean intersection over union of the proposed model can reach 95.21%and 91.03%respectively.The proposed model has realized the mobile supervision through the WeChat enterprise account,and can provide new technical solutions for water conservancy related departments to carry out the intelligent monitoring and timely early warning of floating debris in reservoirs.