Gauss-guided patch-based deep convolutional neural networks for no-reference image quality assessment

Abstract

Patch-based deep convolutional neural network (DCNN) has been proved to have advanced performance in no-reference image quality assessment (NR-IQA). However, these methods generally take global quality score as the quality score of each patch mainly since local quality score is not provided. Unfortunately, the perceived quality of image patch is difficult to maintain a high degree of consistency. Thus, the use of the same global quality score in different patches of the same image may hinder training of DCNNs. In this paper, we propose a universal and nearly cost-free model called Gaussian Random Jitter (GRJ). According to the uncertainty of the perceived quality, GRJ divided the training images into high-confidence distorted images and low-confidence distorted images, and reasonably assigned different local quality scores to each patch through specific gaussian functions with the global quality score as the mean value and the undetermined hyperparameter as the standard deviation. We took one of the most advanced patch-based DCNNs models as backbone and tested the improved performance over three widely used image quality databases. We show that our model can further improve the performance of patch-based models and even help them comparable with those of state-of-the-art NR-IQA algorithms.