Weakly-Correlated Knowledge Distillation for Blind Image Quality Assessment

Current blind image quality assessment (BIQA) techniques have seen significant improvements with the adoption of transfer learning and multi-task learning, which leverage auxiliary knowledge from tasks strongly-correlated with BIQA. However, these approaches typically require large amounts of annotated data with multi-task labels, extensive pre-training, and powerful hardware. In contrast, this paper aims to improve BIQA by transferring knowledge from existing pre-trained high-level vision models that are weakly-correlated with BIQA, without requiring additional pre-training or extra labels. High-level vision models, such as well-developed semantic segmentation models, are sensitive to semantic content but insensitive to image quality, while BIQA models focus on the opposite. Furthermore, a given distortion may result in different quality degradations on different semantic content. To this end, we propose a weakly-correlated knowledge distillation (WCKD) framework for BIQA by transferring knowledge from the semantic segmentation model to the BIQA model. Specifically, we use a pre-trained segmentation model with the same backbone as the BIQA model and propose a correlation constraint that maximizes the feature map distance between the segmentation and BIQA models. This encourages the BIQA model to focus on semantically-irrelevant information which is considered to be quality-related. Additionally, we fuse features from both models to account for varying effects of distortions on different semantic content, improving the accuracy of quality assessment. Experimental results on state-of-the-art BIQA models and datasets demonstrate the superiority of our approach. The source code will be available at:https://shishuntian.github.io/.