基于改进DeeplabV3+的水面多类型漂浮物分割方法研究  

Research on segmentation method of multiple types of floating objects on water surface based on improved DeeplabV3+

作　　者：包学才[1,2] 刘飞燕 聂菊根[1,2] 许小华 柯华盛 BAO Xuecai;LIU Feiyan;NIE Jugen;XU Xiaohua;KE Huasheng(Key Laboratory of Cooperative Sensing and Intelligent Processing of Water Information in Jiangxi Province,Nanchang Institute of Technology,Nanchang 330099,Jiangxi,China;College of Information Engineerin,Nanchang Institute of Technology,Nanchang 330099,Jiangxi,China;Jiangxi Institute of Water Resources,Nanchang,Jiangxi,Nanchang 330099,Jiangxi,China)

机构地区：[1]南昌工程学院江西省水信息协同感知与智能处理重点实验室,江西南昌330099 [2]南昌工程学院信息工程学院,江西南昌330099 [3]江西省水利科学院,江西南昌330029

出　　处：《水利水电技术（中英文）》2024年第4期163-175,共13页Water Resources and Hydropower Engineering

基　　金：国家自然科学基金项目(61961026);江西省水利厅科技项目(202223YBKT19);江西省科技厅重大科技研发专项“揭榜挂帅”制项目(20213AAG01012)。

摘　　要：【目的】为解决传统图像处理方法鲁棒性差、常用深度学习检测方法无法准确识别大片漂浮物的边界等问题,【方法】提出一种基于改进DeeplabV3+的水面多类型漂浮物识别的语义分割方法,提高水面漂浮的识别能力。对所收集实际水面漂浮物进行分类,采用自制数据集进行对比试验。算法选择xception网络作为主干网络以获得初步漂浮物特征,在加强特征提取网络部分引入注意力机制以强调有效特征信息,在后处理阶段加入全连接条件随机场模型,将单个像素点的局部信息与全局语义信息融合。【结果】对比图像分割性能指标,改进后的算法mPA(Mean Pixel Accuracy)提升了5.73%,mIOU(Mean Intersection Over Union)提升了4.37%。【结论】相比于其他算法模型,改进后的DeeplabV3+算法对漂浮物特征的获取能力更强,同时能获得丰富的细节信息以更精准地识别多类型水面漂浮物的边界与较难分类的漂浮物,在对多个水库场景测试后满足实际水域环境中漂浮物检测的需求。[Objective]In order to solve the problems of poor robustness of traditional image processing method and the inability of commonly used deep learning detection method to accurately recognize the boundaries of large floating objects,[Methods]a semantic segmentation method based on the improvement of DeeplabV3+ for the recognition of multiple types of water surface floats was proposed,which improves the recognition ability of water surface floats.By classifying the collected actual water surface floats into categories,a homemade dataset is used for comparison experiments.The algorithm selects the xception network as the backbone network to obtain preliminary float features,introduces the attention mechanism in the enhanced feature extraction network part to emphasize the effective feature information,and incorporates the fully-connected conditional random field model in the post-processing stage to fuse the local information of a single pixel with the global semantic information.[Results]Comparing the image segmentation performance metrics,the improved algorithm mPA(Mean Pixel Accuracy) is improved by 5.73% and mIOU(Mean Intersection Over Union) is improved by 4.37%.[Conclusion]Compared with other algorithmic models,the improved DeeplabV3+ algorithm is more capable of acquiring features of floating objects,and at the same time obtains rich detail information to more accurately identify the boundaries of multiple types of water surface floating objects and more difficult to classify floating objects,which meets the needs of floating object detection in the actual water environment after testing on multiple reservoir scenarios.

关 键 词：深度学习 语义分割 特征提取 漂浮物识别 注意力机制 全连接条件随机场 算法模型 影响因素 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]