A well-perceived, blind image quality assessment algorithm using an enhanced noise feature criterion

Many deep learning-based blind image quality assessment (BIQA) methods achieve high accuracy but rely heavily on complex network architectures and large datasets, which limit their applicability. This study proposes an enhanced perception-based no-reference (NR) BIQA method that incorporates a revised noise feature criterion for immediate and practical use. This approach was motivated by observations that conventional noise feature analysis becomes unstable in images with strong horizontal structures, such as fence-like patterns. To address this limitation, improved noise weighting and decision criteria were introduced. The method was evaluated on four publicly available databases (LIVE, CSIQ, TID2013, and KADID-10k), demonstrating higher or comparable prediction performance relative to the baseline algorithm, as measured by Spearman rank order correlation coefficient (SROCC) and Pearson linear correlation coefficient (PLCC). A detailed comparative analysis of quality estimation performance was conducted between the reference algorithm and the proposed algorithm. The estimated image quality scores were presented side by side, demonstrating that the proposed algorithm achieved more accurate estimations for the 24 perfect and distortion-free images in the TID2013 dataset. The results showed that the proposed algorithm placed all images closer to the 'Excellent' quality region according to Matlab help center description, aligning more closely with the expected evaluation goals than the reference algorithm.