A Ranking-Based Self-Supervised Learning Training Method for No-Reference Image Quality Assessment

In no-reference image quality assessment, the constructed deep neural network models directly rate the image viewing quality without requiring any reference information. However, due to the specificity of this task, the standard datasets are small, and training deep models using traditional methods leads to overfitting; thus, these methods do not have broad real-life applications. Currently, most scholars mitigate this problem by introducing the idea of self-supervision, where models are first pretrained on automatically generated large-scale datasets and later fine-tuned with mean square error (MSE) loss on a specific dataset. However, this approach trains the model to fit the quality labels of only a single image, ignoring the relative quality relationships between images. Since the relative quality relationship between images significantly impacts the generalization performance of the model, we propose a model training method based on ranking self-supervised learning, through which the training model extracts the relative quality features between images and improves its generalization ability. Specifically, we construct two ranking losses for pretraining and fine-tuning, using pairwise ranking to train the model to rank pictures based on the perceptual quality of the images. Finally, we conduct extensive experiments on the proposed method using three state-of-the-art quality evaluation models on seven publicly available datasets. The experimental results show that the proposed training methods significantly improve the prediction accuracy and generalization ability of the resulting models compared to the traditional methods.