A Deep Blind Image Quality Assessment with Visual Importance Based Patch Score

Abstract

Convolutional neural networks (CNNs)-based no-reference image quality assessment (NR-IQA) suffers from insufficient training data. The conventional solution is splitting the training image into patches, assigning each patch the quality score, while the assignment of patch score is not consistent with the human visual system (HVS) well. To address the problem, we propose a patch quality assignment strategy, introducing the weighting map to describe the degree of visual importance of each distorted pixel, integrating the weighting map and the feature map to pool the quality score of each patch. With the patch quality, a CNNs-based NR-IQA model is trained. Experimental results demonstrate that proposed method, named as blind image quality metric with improved patch score (BIQIPS), improves the performance on most of the distortion types, especially on the types of local distortions, and achieves state-of-the-art prediction accuracy among the NR-IQA metrics.