Contrastive Learning in Frequency Domain for Non-I.I.D. Image Classification

Non-I.I.D. image classification is an important research topic for both academic and industrial communities. However, it is a very challenging task, as it violates the famous hypothesis of independent and identically distributed (I.I.D.) in conventional machine learning, and the classifier minimizing empirical errors on training images does not perform well on testing images. In this work, we propose a novel model called Contrastive Learning in Frequency Domain (CLFD) to learn invariant representations for Non-I.I.D. image classification. In CLFD, model learning includes two steps: contrastive learning in the frequency domain for pre-training, and image classification with fine-tuning. In the first pre-training step, anchor, positive and negative images are transformed by Discrete Cosine Transform (DCT) and then projected into vector space. This step is to obtain stable invariant features by minimizing the contrastive loss. In the step of image classification with fine-tuning, the features from ResNet are mapped into the label space by a simple fully connected layer, and the classification loss is utilized to fine-tune the parameters in the ResNet. Extensive experiments conducted on public NICO dataset demonstrate the effectiveness of the proposed CLFD, which outperforms the state-of-the-art methods.