基于对比学习的标签带噪图像分类

标签噪声会极大地降低深度网络模型的性能. 针对这一问题, 本文提出了一种基于对比学习的标签带噪图像分类方法. 该方法包括自适应阈值、对比学习模块和基于类原型的标签去噪模块. 首先采用对比学习最大化一幅图像的两个增强视图的相似度来提取图像鲁棒特征; 接下来通过一种新颖的自适应阈值过滤训练样本, 在模型训练过程中根据各个类别的学习情况动态调整阈值; 然后创新性地引入基于类原型的标签去噪模块, 通过计算样本特征向量与原型向量的相似度更新伪标签, 从而避免标签中噪声的影响; 在公开数据集CIFAR-10、CIFAR-100和真实数据集ANIMAL10上进行对比实验, 实验结果表明, 在人工合成噪声的条件下, 本文方法实验结果均高于常规方法, 通过计算图像鲁棒的特征向量与各个原型向量的相似度更新伪标签的方式, 降低了噪声标签的负面影响, 在一定程度上提高模型的抗噪声能力, 验证了该模型的有效性.

Label Noisy Image Classification Based on Contrastive Learning

Label noise can greatly reduce the performance of deep network models. To address this problem, this study proposes a contrastive learning-based label noisy image classification method. The method includes an adaptive threshold, contrastive learning module, and class prototype-based label denoising module. Firstly, the robust features of the image are extracted by maximizing the similarity between two augmented views of the same image using contrastive learning. Then, a novel adaptive threshold filtering training sample is used to dynamically adjust the threshold based on the learning status of each class during model training. Finally, a class prototype-based label denoising module is introduced to update pseudo-labels by calculating the similarity between sample feature vectors and prototype vectors, thus avoiding the influence of label noise. Comparative experiments are conducted on the publicly available datasets CIFAR-10 and CIFAR-100 and the real dataset ANIMAL10. The experimental results show that under the condition of artificially synthesized noise, the proposed method outperforms conventional methods. By updating pseudo-labels based on the similarity between the robust feature vector of the image and various prototype vectors, the negative impact of noisy labels is reduced, and the anti-noise ability of the model is improved to certain extent, verifying the effectiveness of the proposed model.