Research On Image Quality Evaluation Algorithm Based On Human Visual System

The objective evaluation of image quality is very important for image processing system such as image transmission,compression and other technologies.The objective image evaluation includes full-reference,semi-reference and no-reference.Among them,full-reference evaluation is the most direct and effective,and no-reference evaluation is the most practical in actual scene.Based on the previous studies,this paper focuses on the feature extraction and evaluation algorithm design related to human visual perception.The main work and innovation in this paper are:(1)The visual forming mechanism and the perceptual characteristics of the human visual system are summarized.This is the basis for designing the image quality evaluation algorithm.Understanding human visual characteristics can better design the evaluation algorithm consistent with human subjective feelings.(2)In the aspect of full-reference,we proposed an evaluation algorithm based on gradient direction selection.Firstly,the gradient of the image is calculated in two different orthogonal systems,and the larger gradient amplitude is selected by pixel in the two gradient maps.The gradient selection of the distorted image is based on the reference image,and the gradient magnitudes from the same orthogonal system are selected by pixel.Thereby the gradient similarity between the reference and distorted image is calculated.In order to evaluate the color images,the image is converted from RGB spatial to the color space which the luminance and chrominance are separated.Then the gradient similarity is calculated on one luminance channel and two chrominance channels,respectively.Finally,by sampling the image to different scales to simulate the actual observation scene,different weights are given to the scales to obtain the final perceived quality score.(3)In the aspect of no-reference,we proposed a method based on the combination of natural scene statistical model and LBP histogram feature.First,the Generalized Gaussian Distribution fitting parameters are used as the features of the natural scene statistics,and then the relevant features are extracted based on the texture information of the image.The rotation invariant LBP map is calculated based on the gradient map,and then the dark channel is utilized as the weight to obtain the histogram,which is the second set of features.Then the original image is convoluted using a low pass filter,and the features are repeatedly calculated on the filtered image to form the final image sensing feature.Finally,the mapping of features and subjective scores is established by support vector regression.The experiments presented in this paper verity the proposed algorithm has higher prediction accuracy and robustness,and the comprehensive performance is better.(4)This paper designed and implemented an image quality evaluation system.This system has two functional modules,which implement the two evaluation algorithms proposed above respectively.The system implements batch processing in no-reference quality assessment,and shows the results clearly.