Multi-scale Transformer with Decoder for Image Quality Assessment

Blind image quality assessment (BIQA) is of great significance in image processing field. However, due to diverse image content and complex types of distortions, the issue of BIQA has not been fully resolved. To address this issue more effectively, in this paper, we propose a framework based on Vision Transformer for BIQA called MSIQT. This model aims to extract image features more effectively and achieve a more accurate representation of quality. Specifically, at the input end, we adopt a multi-scale input approach to enrich the image features and utilize ResNet-50 for feature extraction. At the output end, a decoder is introduced to interpret quality-aware vectors obtained from image features. Experiments on four image quality assessment datasets prove that the proposed method outperforms or is comparable to state-of-the-art approaches.