面向类别不平衡语义分割的损失函数的研究  

Study on Loss Functions for Semantic Segmentation Based on Imbalanced Datasets

作　　者：孙盛[1] 雷松 徐志佳[2] 胡忠文[3] SU Sheng;LEI Song;XU Zhi-jia;HU Zhong-wen(School of Computers,Guangdong University of technology,Guangzhou Guangdong 510006,China;School of mechanical engineering,Guiyang University,Guiyang Guizhou 550005,China;Key Laboratory of Geo environmental Monitoring in the Greater Bay Area,Ministry of Natural Resources,Shenzhen University,Shenzhen Guangdong 518000,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006 [2]贵阳学院机械工程学院,贵州贵阳550005 [3]深圳大学自然资源部大湾区地理环境监测重点实验室,广东深圳518000

出　　处：《计算机仿真》2024年第5期310-317,324,共9页Computer Simulation

基　　金：国家自然科学基金项目(61672007);广东省国际合作领域项目(2019A050509009);广东省海洋与渔业厅渔港建设和渔业发展专项(A201701D04);自然资源部大湾区地理环境监测重点实验室开放基金(2019002)。

摘　　要：在图像语义分割的任务中,数据集的类别不平衡分布是一种普遍存在的现象;通过构造更加有效的损失函数来减轻其带来的不利影响是具有较大研究价值的。以损失函数为研究对象,利用训练样本像素的空间分布特性和相关性,改进了Focal损失函数的尺度度量方法,提出了一个基于迭代批次的权重自适应调整的Adapt-Focal损失函数。以类别不平衡的SAR图像数据集GDUT-Nansha为实验对象,采用三个经典的卷积神经网络DeepLabV3+、U-Net和RWL-ENet完成了定量实验。实验结果表明,所提出的Adapt-Focal损失函数相比于CE、FL、dfl、acw、lovász和softiou损失函数,有效提高了少样本地物类别裸地和道路的分割精度IoU和PA值;同时,整体分割精度指标mPA和mIoU均得到较大幅度的提高。验证了Adapt-Focal损失函数在类别不平衡图像语义分割中的有效性。In the task of image semantic segmentation,the class imbalance distribution of the dataset is a common phenomenon;Constructing a more effective loss function to mitigate its adverse effects is of great research value.This paper takes the loss function as the research object,improves the scale measure of the Focal loss function using the spatial distribution characteristics and correlation of the training sample pixels,and proposes an Adapt-Focal loss function with adaptive adjustment of weights based on iterative batches.Quantitative experiments were completed using three classical convolutional neural networks,DeepLabV3+,U-Net and RWL-ENet,with the class imbalance SAR image dataset GDUT-Nansha as the experimental object.The experimental results show that the proposed Adapt-Focal loss function effectively improves the segmentation accuracy IoU and PA values for the few-sample feature class bare ground and roads compared with the CE,FL,dfl,acw,lovász and softiou loss functions;meanwhile,the overall segmentation accuracy metrics mPA and mIoU are both substantially improved.The effectiveness of Adapt-Focal loss function in semantic segmentation of category imbalanced images is verified.

关 键 词：合成孔径雷达图像 语义分割 类别不平衡 损失函数 权重自适应调整 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]