No reference image quality assessment based on axis attention Transformer convolutional neural network
ZHAO Haiyang
WANG Changzhong
LYU Xiang
SHAO Mingwen
Abstract：No-reference image quality assessment(NRIQA)has always been an important research direction in the fields of image processing and computer vision,with the goal of estimating image quality through objective assessment.In recent years,no-reference image quality assessment methods based on convolutional neural network(CNN)and Transformers have received extensive attention.However,due to the complex structure of the Transformer,most existing algorithms have the problem of high computational complexity.To this end,this paper proposes a multi-level convolutional neural network based on Transformer(TBMCNN),which only extracts features from the horizontal and vertical axis of the Transformer,significantly improving the network performance.The multi-level convolutional neural network based on Transformer consists of four core modules:the shallow feature extraction module(SFEM),the parameter sharing module(PSM),the multi-level Transformer feature extraction module(MTFEM),and the fully connected network module(FCNM).Among them,the multi-level Transformer feature extraction module is designed with parallel branches,which are respectively used for adaptive multi-scale convolution and axis attention Transformer.This not only enables more effective extraction and fusion of features but also significantly enhances the computing speed of the network.The experimental results show that the multi-level convolutional neural network based on Transformer has better performance and generalization ability without consuming excessive resources.
Keywords：no-reference image quality assessment(NRIQA)convolutional neural network(CNN)axis attention Transformermulti-scale convolution
Publication Date：2025-12-15
Online Publishing Date：2026-04-01(First online date of this platform, not the publication date of the document)
Pages：16( 273-288 )