Contrastive learning based deep neural network for no-reference image quality assessment

Abstract

In this work, we propose contrastive learning based deep neural network model for no-reference image quality assessment (NR-IQA). The proposed deep neural network model consists of 11convolutional neural networks (CNN) in parallel, each CNN initializing the weight to extract the characteristics of a specific group. For the initialization of weights of 11 CNNs, 11 groups are generated from LIVE dataset. The 11 groups are divided based on the distortion types and subjective image quality scores of the dataset, and the data in each group do not need to be mutually exclusive. With 11 groups generated from the dataset, 11 siamese networks that extract the characteristics of each group are pre-trained through contrastive learning. We then fine-tune the entire network on the target subjective image quality score. This method shows better performance than the previous NR-IQA model by using only limited image quality assessment data and suggests a feature extraction method that is meaningful for image quality assessment.