A Distortion Aware Image Quality Assessment Model

Recent image quality metrics have taken advantage of large pre-trained vision models. However, these models require a large amount of data for training and/or have a high number of fine-tuned parameters. In addition, in such metrics, the local features are often ignored in the quality regression which can decrease the accuracy of the approach. In this study, we propose a new image quality metric that is focused on using distortion classification to obtain distortion-aware features and improve the performance of our image quality metric. We first extract local distortion features from a distortion classification task and then combine them with global content-related information to create quality features. Then, the features are aggregated as quality features, which are injected into a multilayer perceptron regressor to predict the image quality. Experiments on six common subjective datasets show that the proposed model achieves competitive performance while using a drastically lower number of parameters compared to the current state-of-the-art image quality metrics.