This paper presents a no-reference image quality assessment metric that makes use of the wavelet subband statistics to evaluate the levels of distortions of wavelet-compressed images. The work is based on the fact that for distorted images the correlation coefficients of the adjacent scale subbands change proportionally with respect to the distortion of a compressed image. Subband similarity is used in this work to measure the correlations of the adjacent scale subbands of the same wavelet orientations. The higher the image quality is (i.e., less distortion), the greater the cosine similarity coefficient will be. Statistical analysis is applied to analyze the performance of the metric by evaluating the relationship between the human subjective assessment scores and the subband cosine similarities. Experimental results show that the proposed blind method for the quality assessment of wavelet-compressed images has sufficient prediction accuracy (high Pearson Correlation Coefficient, PCCs), sufficient prediction monotonicity (high Spearman Correlation Coefficient SCCs) and sufficient prediction consistency (low outlier ratios) and less running time. It is simple to calculate, has a clear physical meaning, and has a stable performance for the four image databases on which the method was tested.
Comparison of no-reference image quality assessment machine learning-based algorithms on compressed images
