Full-reference image quality assessment based on LS convolution

Abstract

This study proposes a full-reference image quality assessment (FR-IQA) algorithm, named LSCNN, which conforms to the “large perception, small aggregation” characteristic of the human visual system. The LSCNN algorithm extracts important features that conform to human subjective perception by replacing some convolutional layers of the convolutional neural network (CNN) with LSConv (Large-Small Convolution), and uses two layers of fully connected layers for regression prediction. LSConv can achieve global context capture and local feature fusion to optimize convolution efficiency, simulate the feature extraction process of the human visual system, and make the objective prediction of image quality evaluation results closer to the level of human subjective evaluation. The LSCNN algorithm was evaluated on the LIVE, TID2013, CSIQ, SCID and SIQAD datasets, and the experimental results showed that its performance was superior to some existing FRIQA algorithms.