In real-time applications such as Digital Imaging and COmmunications in Medicine(DICOM), social media applications and video conferencing images are subjected to distortion at stages like digitization, compression, and transmission. Distortions can degrade the human visual experience. Hence it is crucial to design an efficient perceptual quality estimation model. A simple yet efficient HITS-NRIQA method is proposed to improve the patch sampling strategy to achieve accurate and robust deep learning for the BIQA tasks. In contrast to the previously published work, this paper has exploited the advantage of more informative patches along with adaptive weighting to effectively learn the image quality. Moreover, the proposed approach retains simplicity while significantly improving its accuracy over state of the art on BIQA task and can be equipped with other tasks like image restoration, compression, etc.

Quality degradation due to the compression and the transmission of images is a significant threat to multimedia applications. Blind image quality assessment (BIQA) is a principal technique to measure the distortion and dynamically set the optimal parameters for developing image compression standards, image restoration algorithms, etc. Insufficient training data with quality scores are a challenge for image quality assessment (IQA) tasks. The existing solutions to deep-convolution-neural-network-based BIQA, which rely on patchwise training, struggle to find an ideal set of patches consistent with the human visual system. To address these issues, HIerarchical paTch Selection (HITS) is proposed. HITS keeps the patches with considerable details in each quadrant according to their intensity variance scatter ratio (IVSR). IVSR identifies the best nonhomogeneous patches by computing patch intensity variance. Extensive trials are conducted, and the performance reveals that the proposed approac...