文献类型：期刊文献

中文题名：基于类别不平衡数据集的图像实例分割方法

英文题名：Image Instance Segmentation Method Based on Class-imbalanced Dataset

作者：范馨月[1,2];鲍泓[1,2];潘卫国[1,2]

第一作者：范馨月

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100027

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2022

卷号：48

期号：12

起止页码：224-231

中文期刊名：计算机工程

外文期刊名：Computer Engineering

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家自然科学基金(61932012,62102033)。

语种：中文

中文关键词：长尾分布;实例分割;数据增强;损失函数;深度学习

外文关键词：long tail distribution;instance segmentation;data augmentation;loss function;deep learning

摘要：随着深度学习在计算机视觉领域取得重大进展,包含多种类别的数据集不断被提出,但由于自然采集的数据集往往存在类别不平衡并呈现长尾分布的情况,导致稀有类的数据特征被频繁类的数据特征所抑制,从而严重影响模型的检测性能。为解决上述问题,提出一种新的图像实例分割方法。采用长尾实例分割数据集进行研究实验,使用基于目标尺度的数据增广方法对数据集进行处理,以达到扩充训练样本的目的,增加稀有类的目标数量,同时对稀有类数据进行重采样,解决稀有类的类别数据量过小的问题,提升模型在长尾数据集的鲁棒性。在此基础上,将均等化损失函数融入Mask R-CNN实例分割网络,以降低频繁类的数据特征对稀有类数据特征的抑制性。实验结果表明,该方法在LVIS实例分割数据集上的检测精度提升了4.9%,达到了25.7%,同时AP_(r)、AP_(c)、AP_(f)分别达到了16.2%、26.1%、30.4%,相比Baseline方法均有明显提升,在消融实验上的结果也表明该方法能有效解决长尾类问题。
With the significant progress made by deep learning in the field of computer vision,datasets containing various categories are constantly being proposed.However,in naturally collected datasets,there are often classimbalances and long-tailed distributions.As a result,the data features of rare classes are suppressed by the data features of frequent classes,seriously affecting the performance of the model.Therefore,the algorithm cannot be implemented well.To solve this problem,this paper proposes a method for image instance segmentation in which the long-tailed instance segmentation dataset is used for experiments.First,the dataset is processed by the data augmentation method based on the target scale to expand the training samples and increase the target number of rare classes.Rare class data are resampled to solve the problem of the excessively small number of rare class data,thereby improving the robustness of the model in long-tailed datasets.Finally,the equalization loss function is integrated into the Mask Region-based Convolutional Neural Network(R-CNN)instance segmentation network to reduce the inhibition of the data features of the frequent class to the data features of the rare class.In an experimental verification on the Large Vocabulary Instance Segmentation(LVIS)dataset,the method proposed improves the detection accuracy by 4.9%,reaching 25.7%.In addition,AP_(r),AP_(c),and AP_(f) reach 16.2%,26.1%,and 30.4%,respectively,which are significantly improved compared with the Baseline method.The results of ablation experiments performed using the proposed method show that it can solve the long-tailed-distribution problem.