Blind Quality Evaluation for Color Image Based on Naturalness and Structure Distortions

Based on the principle that the grayscale value of an image is obtained through a linear calculation of the R, G, and B color components, this study proposes a blind quality assessment method for color images that accounts for both naturalness and structural distortions. The method extracts features from the three color channels and employs a neural network to train a quality prediction model. First, through pyramid decomposition and split normalization, the three color channels are transformed into the wavelet domain, where quality-aware features(QAFs) are extracted based on natural scene statistics to measure naturalness distortion. Second, QAFs are extracted from the gradient domain of the three color channels to measure structural distortion. Additionally, the moments and two-dimensional entropy of the three color channels are incorporated as supplementary features. Finally, a quality prediction model is constructed using a backpropagation neural network trained on the extracted features. Experimental results on three color gamut-mapped image databases with color distortion, including Basic Study, Image Gamut, and Local Contrast, demonstrate that the proposed method achieves superior performance in assessing color distortion.