Accuracy and generalization improvement for image quality assessment of authentic distortion by semi-supervised learning

Abstract

Image quality assessment of authentically distorted images constitutes a indispensable part of numerous computer vision tasks. Despite the substantial progress in recent years, accuracy and generalization performance is still unsatisfactory. These challenges are primarily attributed to the scarcity of labeled images. In order to increase the amount of images for training, we use semi-supervised learning to combine labeled images and specifically selected unlabeled images. In our new training paradigm, denominated Selected Data Retrain under Regularization, the selection criteria of unlabeled images is based on the supposition that an image and a certain of its patches ought to have approximate image quality scores. Unlabeled images that meets the aforementioned criteria, named as Highly Credible Unlabeled Images, mitigate the problem of scarcity, thus, improve accuracy. However generalization may be compromised due to selection procedure’s reliance on labeled images and presence of coherent variance existed between labeled images and unlabeled images. Therefore we incorporate a sorting loss function to reduce variation within the new dataset of labeled images and specifically selected unlabeled images, and thus achieve better generalization. The effectiveness of our proposed paradigm is empirically validated using public datasets. Codes are available at https://github.com/dvstter/SDRR_IQA.