Blind quality evaluation for tone-mapped images by exploiting statistical characteristics and deep perceptual features

The tone-mapping operator (TMO) aims to convert high dynamic range images into low dynamic range images, enabling them to be displayed on standard monitors. However, this conversion process inevitably leads to a decrease in image quality, such as structural damage, contrast degradation and color artifacts, which impacts human’s visual perception. Evaluating the visual quality of tone-mapped images (TMIs) effectively remains a challenge. To address this issue, this paper proposes a novel blind metric for evaluating the quality of TMIs based on statistical characteristics and deep perceptual features. Specifically, for statistical features, we firstly extract quality-aware features that capture structure-texture variations from local fractal dimension maps and first digit distribution statistics from wavelet and gradient domains to capture structural changes in TMIs. Additionally, considering that the TMO procedure introduces noticeable color artifacts, we further extract color statistical features to capture color degradation. For deep perceptual features, a pre-trained CNN model is utilized to extract deep learning-based features that represent semantic changes in TMIs. Finally, considering the significant difference between the extracted statistical and deep perceptual features, we leverage an ensemble learning scheme for feature training and quality score generation. Experimental results conducted on three publicly available benchmark databases demonstrate that our method achieves better performance than existing traditional 2D quality metrics and specifically-designed TMI quality metrics.