摘要
长尾识别是计算机视觉领域最具挑战性的问题之一。在现实世界中长尾识别具有广泛的应用,研究长尾识别具有重要意义。对于长尾分布数据来说,由于类与类之间样本量不平衡,以及占比众多的尾部类缺少足够的训练样本,使其在训练过程中很难找到各类间的明确界限。为解决这一问题,将元预训练和监督对比学习结合起来,提出了基于平衡对比学习策略的长尾识别方法MBCP-BB(meta balanced contrastive pre-training and batch balance)。MBCP-BB采用解耦学习方式进行模型训练:通过预训练获得具有优异特征表示能力的特征提取器,在微调阶段,固定特征提取器,重新训练分类器。该方法突出特征学习的重要性,设计了平衡对比学习策略指导特征学习过程,从而使监督对比学习技术能有效应用于长尾识别场景。进行特征学习时,首先适当减少头部类样本,并利用少样本图像生成技术为尾部类生成新样本;之后以每类的类原型作为补充样本用于训练。解耦学习训练模式下,充分挖掘了特征提取器与分类器的潜力,在增强模型特征学习能力的同时,大大简化了分类器的训练过程。在几个长尾基准数据集上进行了大量实验,并与7个代表性的算法从多个角度进行了实验比较,实验结果表明该方法优于比较的算法。
Long-tailed recognition is one of the most challenging problems in computer vision.Long-tailed recognition has a wide range of applications in the real world,and it is of great significance to study long-tailed recognition.For long-tailed distribution data,due to the unbalanced sample size between classes and the lack of sufficient training samples for the large tail classes,it is difficult to find a clear boundary between classes during the training process.To address this issue,we combine meta pre-training and supervised contrastive learning,and propose MBCP-BB(meta balanced contrastive pre-training and batch balance),a long-tailed recognition method based on a balanced contrastive learning strategy.MBCP-BB adopts a decoupled learning method for model training:A feature extractor with excellent feature representation ability is obtained through pre-training,and in the fine-tuning stage,the feature extractor is fixed and the classifier is retrained.This method highlights the importance of feature learning,and designs a balanced contrastive learning strategy to guide the feature learning process,so that supervised contrastive learning techniques can be effectively applied to long-tailed recognition scenarios.When performing feature learning,first reduce the samples of the head classes appropriately,and use the few-shot image generation technology to generate new samples for the tail classes;then use the class prototype of each class as supplementary samples for training.In the decoupled learning training mode,the potential of the feature extractor and classifier is fully mined,and the training process of the classifier is greatly simplified while enhancing the feature learning ability of the model.A large number of experiments are carried out on several long-tailed benchmark datasets,and compared with seven representative algorithms from multiple perspectives,the experimental results show that the proposed method is superior to the compared algorithms.
作者
孔令权
翟俊海
KONG Lingquan;ZHAI Junhai(College of Mathematics and Information Science,Hebei University,Baoding 071002,China;Key Laboratory of Machine Learning and Computational Intelligence,Hebei University,Baoding 071002,China)
出处
《西北大学学报（自然科学版）》
CAS
CSCD
北大核心
2024年第4期677-688,共12页
Journal of Northwest University（Natural Science Edition）
基金
河北省科技计划重点研发项目(19210310D)
河北省自然科学基金(F2017201026)。
关键词
长尾识别
元学习
预训练
监督对比学习
批次平衡训练
long-tailed recognition
meta-learning
pre-training
supervised contrastive learning
batch balance training
