基于受控正样本采样与平衡负样本挖掘的无监督图像分类算法OA

Contrastive learning often suffered from false positive samples and false negative samples during positive sample augmentation and negative sample mining,which could lead to degraded performance of unsupervised image classification algorithms.To address this problem,a controlled positive sample sampling and balanced negatives sample mining(CPS-BN)algorithm for unsupervised image classification was proposed.Specifically,the CPS algorithm was designed to generate semantically consistent yet moderately diverse positive sample pairs through controlled view sampling,thereby reducing incorrect positive sample matching in fine-grained scenarios.Meanwhile,the BN algorithm was formulated to construct a negative sample selection probability based on similarity differences,and ineffective samples as well as false negative samples were adaptively filtered to enhance effective gradients and stabilize training.The results on the Canadian Institute for Advanced Research-10(CIFAR-10)image dataset showed that the clustering accuracy(ACC)of the CPS-BN algorithm was improved by 4.1 percentage points compared with the strongly augmented contrastive clustering(SACC)algorithm.On the ImageNet fine-grained dog dataset,the ACC of the proposed algorithm was further improved by 3.8 percentage points over the SACC algorithm.The CPS-BN algorithm effectively alleviated sample selection bias in contrastive learning and improved the quality of feature representation and clustering performance,which was well suited for unsupervised image classification scenarios.