基于双边采样对比增强的不平衡图像分类研究  

UNBALANCED IMAGE CLASSIFICATION BASED ON BILATERAL SAMPLING CONTRAST ENHANCEMENT

作　　者：彭祎祺 林珊玲 林志贤[1,2] 周雄图 郭太良[1,2] Peng Yiqi;Lin Shanling;Lin Zhixian;Zhou Xiongtu;Guo Tailiang(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350116,Fujian,China;Fujian Science&Technology Innovation Laboratory for Optoelectronic Information of China,Fuzhou 350116,Fujian,China;School of Advanced Manufacturing,Fuzhou University,Jinjiang 362200,Fujian,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350116 [2]中国福建光电信息科学与技术创新实验室,福建福州350116 [3]福州大学先进制造学院,福建晋江362200

出　　处：《计算机应用与软件》2023年第6期205-210,共6页Computer Applications and Software

基　　金：国家重点研发计划项目(2016YFB0401503);广东省科技重大专项项目(2016B090906001);福建省自然科学基金项目(2020J014687)。

摘　　要：针对训练数据不平衡问题,提出一种基于双边采样对比增强的渐进式学习分类网络(BBRCL)。该方法对训练数据分别进行均匀采样和倒置采样,将均匀采样后的数据使用随机图像增强生成两个视图;把视图样本和倒置采样样本分别送入卷积神经网络进行特征学习,进一步利用对比学习策略得到高质量视图样本特征;将提取到的两种特征通过累加学习融合训练分类器,实现长尾数据分类。在两个不平衡数据集上进行实验对比,结果表明,该模型能在不损害学习深度特征能力的情况下有效处理数据不平衡问题,且分类准确率相对传统分类方法提升6%。To solve the problem of unbalanced training data,a progressive learning classification network based on bilateral sampling contrast enhancement is proposed,which names BBRCL.This method performed uniform sampling and reversed sampling respectively,and the sample data were enhanced by using random images to generate two views.The generated views samples and reversed samples were put into convolution neural network for feature learning respectively,further more used contrastive learning strategy to obtain high-quality view sample features.The extracted feature vectors were used to accumulative learning unit for training classifier to realize long-tail data classification.The experimental comparison between the two unbalanced data sets shows that the proposed method can effectively deal with the unbalanced data problem without damaging the ability of depth characteristics learning,and the classification accuracy is increased by 6%compared with the traditional classification methods.

关 键 词：长尾数据 数据采样 特征学习 对比学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]