Deep Learning-Based Non-Reference Image Quality Assessment Using Vision Transformer with Multiscale Dual Branch Fusion

Abstract

Non-Reference image quality assessment does not rely on reference images, so it is not easy to directly obtain the actual label of image quality. Current datasets are often limited in scale, and the labeling process is highly subjective, resulting in limited consistency and accuracy in evaluation results. This study focuses on the research of reference-free image quality evaluation based on Vision Transformer multi-scale dual-branch fusion, aiming to build an intelligent system that can accurately and quickly evaluate image quality without original image reference through deep learning technology. In this study, the Vision Transformer model, combined with a multi-scale dual-branch fusion strategy, is used to conduct an in-depth exploration of quality assessment in complex image scenes. The experimental results show that the evaluation accuracy of the system on large-scale image data sets reaches 94%, and the processing speed is 30% higher than that of the traditional method, which is significantly better than the 75% accuracy and lower processing efficiency of the conventional algorithm.