Multi-dataset learning with channel modulation loss for blind image quality assessment

In Blind Image Quality Assessment (BIQA), due to the problem of laborious labeling, it is perceived as the intractability of collecting a new large-scale dataset that has plentiful images with a large diversity in distortion and scene. Therefore, to develop a general model, training with data from diverse datasets could be a viable solution and hold significant value. A straightforward solution is to mix multiple datasets to train a robust model. However, as IQA datasets vary in terms of distortion types and labeling mechanisms, models fail to adapt well across various datasets. To solve the multi-dataset adaption problem, we propose a Channel Modulation Loss that encourages each channel to adaptively enhance/diminish dataset-corresponding features. In detail, when training a CNN with data from all the datasets, certain channels acquire knowledge that can be applied across some datasets, whereas others are only effective for a specific dataset. Therefore, we propose a dataset-specific kernel as supervision, by which channel attention can be explicitly optimized to emphasize the channels corresponding to the dataset of an input image. Finally, extensive experiments on six IQA databases show the promise of the learned method in blindly assessing image quality.