Generative Contrastive Learning for Multi-Label Image Classification

The multi-label image classification task aims to predict the labels of multiple objects in an image. Existing research has shown that modeling the frequent co-occurrence relationships between objects can improve model classification performance. However, most current models overemphasize these co-occurrence relationships while neglecting the intrinsic information within the image, resulting in a reduced ability to distinguish between different objects. In this paper, we propose a multi-label image classification method based on contrastive learning to maintain the model's sensitivity to different objects. Traditional contrastive learning methods for multi-label tasks often perform poorly due to the lack of sufficient contrastive samples. To address this, we introduce a generative contrastive learning method, which generates a large number of reliable contrastive samples to enhance the effectiveness of contrastive learning. Our experiments on the MSCOCO2014 dataset show that the proposed method achieves state-of-the-art classification performance.