Compressed-Domain Vision Transformer for Image Classification

Compressed-domain visual task schemes, where visual processing or computer vision are directly performed on the compressed-domain representations, were shown to achieve a higher computational efficiency during training and deployment by avoiding the need to decode the compressed visual information while resulting in a competitive or even better performance as compared to corresponding spatial-domain visual tasks. This work is concerned with learning-based compressed-domain image classification, where the image classification is performed directly on compressed-domain representations, also known as latent representations, that are obtained using a learning-based visual encoder. In this paper, a compressed-domain Vision Transformer (cViT) is proposed to perform image classification in the learning-based compressed-domain. For this purpose, the Vision Transformer (ViT) architecture is adopted and modified to perform classification directly in the compressed-domain. As part of this work, a novel feature patch embedding is introduced leveraging the within- and cross-channel information in the compressed-domain. Also, an adaptation training strategy is designed to adopt the weights from the pre-trained spatial-domain ViT and adapt these to the compressed-domain classification task. Furthermore, the pre-trained ViT weights are utilized through interpolation for position embedding initialization to further improve the performance of cViT. The experimental results show that the proposed cViT outperforms the existing compressed-domain classification networks in terms of Top-1 and Top-5 classification accuracies. Moreover, the proposed cViT can yield competitive classification accuracies with a significantly higher computational efficiency as compared to pixel-domain approaches.