Learning from Rankings with Multi-level Features for No-Reference Image Quality Assessment

Deep neural networks for image quality assessment have been suffering from a lack of training data for a long time, as it is expensive and laborious to collect sufficient subjective mean opinion scores (MOS). The Siamese network (learning from rankings) makes it possible to use images with only rough labels, such as the relative quality. On the other hand, features from intermediate layers, which possess local information highly sensitive to the quality degradation, are overlooked in the direct use of DNNs. In light of these findings, in this paper, we propose a framework for NR-IQA based on transferring learning from the Siamese network to the traditional CNNs by exploiting features from multiple layers. Specifically, we first train a Siamese network on a large artificially generated dataset. Then, we fine-tune the network with a small number of MOS-labeled images to match the perceptual scaling of human beings. In both steps, features from several layers are combined before being fed into the regression layer for the final score. Experimental results are presented, validating the effectiveness of this transfer-learning framework when considering multi-level information. Furthermore, performance comparable to state-of-the-art NR-IQA approaches on standard IQA datasets is achieved.