Blind Quality Assessment of Tone-Mapped Images with Multi-scale Visual Feature Extraction Neural Network

To guarantee the quality of high dynamic range image (HDRI), various tone-mapped operators (TMOs) have been designed to display HDRI on traditional displays recently. Naturally, the image perceptual quality deteriorates seriously due to the inevitable distortions under different TMOs. In this paper, we propose a multi-scale visual feature extraction neural network for blind image quality assessment (BIQA) of TMIs. Specifically, hierarchical image decomposition is elaborately considered to mimic the hierarchical perception mechanism in the human visual system, expecting to better extract and fuse the multi scale features for quality prediction. Besides, under the proposed learning framework, the procedure of feature extraction, multi-scale feature fusion and quality prediction can be jointly optimized in an end-to-end manner. The experiments verify the stable performance of the proposed method on two public TMIs datasets.