Enhancing Blind Image Quality Scores Using Multi-Scale Feature Extraction

The application of BIQA is necessary in cases where no references on images are present, yet the existing techniques are insufficient to catch the whole range of distortions. This failure is especially frequently marked when distortions cover more than one scale. Based on multiscale analyses of image information, recent prospects have been promising for algorithms based on feature extraction being able to break through this limitation and provide a more detailed insight into local and global degradations. This review of the existing literature discusses state-of-the-art multiscale BIQA models in terms of their capabilities to generalize across datasets and how well they work with distortion types. A broad, overall analysis of recent contributions reveals the significant weaknesses of existing techniques that fail to reliably describe complex and mixed distortions and maintain cross dataset robustness. Using Laplacian Pyramid Networks fill these gaps, we present a new multiscale feature extraction framework. Using LPN, an image breaks up into its frequency bands, capturing more details precisely on the higher end and structure on the lower so that quality prediction accuracy attempts are done by improving and enhancing predictions. The current survey gives special attention to the role played by multiscale techniques in BIQA and introduces LPN as a potential solution for quality assessment performance over a large spectrum of visual distortions.