Concurrent Assessment of Image Quality Score and Distortion With Context-Fovea Integrated CNN

This paper proposes a compact framework for simultaneous prediction of image quality scores and corresponding distortion types utilizing a convolutional neural network (CNN) architecture with three integrated streams: a context stream to learn low-resolution features and two fovea streams to learn high-resolution features that operate on the center region. We evaluate the proposed method on the TID2013 and KADID-10k databases by focusing on four prevalent distortions including Gaussian noise, Gaussian blur, JPEG compression, and JPEG2000 compression. Through empirical evaluation, we demonstrate the efficacy of the proposed architecture in accurately predicting quality scores and distortion types. In KADID-10k, the proposed method achieves a distortion classification accuracy of 90.741%, with Pearson correlation coefficients of 94.067% for score prediction and Spearman correlation coefficients of 93.333%. Similarly, in TID2013, the proposed method attains distortion classification accuracy of 92.087%, with Pearson correlation coefficients of 95.062% for score prediction and Spearman correlation coefficients of 96.279%. The experimental results prove the effectiveness of the proposed context and fovea stream CNN architecture for simultaneous IQA and distortion classification.