No-Reference Image Quality Assessment Using FPLBP Features with Laplacian of Gaussian

Abstract

The image quality assessment (IQA) process seeks to determine how people perceive the quality of an image it remains necessary for many image processing and computer vision applications. In this paper, we propose a novel noreference IQA method based on quality-aware texture descriptors constructed using Four-Patch Local Binary Patterns (FPLBPs). Our First, to improve the description of local textures typically affected by visual distortions, we propose modeling them using a multiscale approach. In which, each image is rendered on several subscales using Laplacian-Gaussian (LoG) filters at varying frequency bandwidths. Second, obtained subband images are then used to calculate the FPLBPs descriptors. Second, we calculate the FPLBP descriptors for each of the subband images. In fact, FPLBP consists of modeling the spatial structure based on local intensity variations across each patch, defining local texture representations as local histograms, and generating the global feature vector by concatenating these local histograms. This reflects the overall perceived image quality. Lastly, we train a Support Vector Regressor (SVR) to estimate the Mean Opinion Score (MOS) directly from the feature vectors. Evaluation using a benchmark for IQA, shows that our method has strong associations with subjective quality scores.