Research On Image Quality Assessment Based On Saliency Analysis

Image quality assessment uses mathematical calculation models to measure image quality consistent with subjective assessments.It is a fundamental and important task in the field of computer vision.The image quality assessment is mainly to make the evaluation of the image fit the human visual system,and also to avoid the influence of the bias of the individual's subjective emotions on the evaluation of things.It plays an important role in guiding compression coding,network traffic monitoring,how to choose camera parameters,and processing images.Among them,the image quality assessment is divided into full reference image quality evaluation,semi-reference image quality assessment and no reference quality assessment according to the amount of reference image information.For more than a decade,psychologists,neurobiologists,and computer scientists have conducted extensive and in-depth research on visual salience.In recent years,research scholars have introduced visual saliency features in the direction of image quality assessment to help the algorithm improve performance.Visual saliency reflects the human visual system's attention to objects in the image and helps people extract key information from the image.The full reference image quality assessment algorithm based on visual saliency is more in line with the human visual system's perception of images,so the performance is even better.However,it is difficult to evaluate the image quality by using only the visual saliency feature.In addition to the visual saliency feature,the full reference image quality assessment still combines various basic features to obtain better evaluation performance.In this paper,the visual saliency,gradient magnitude and color similarity are studied in depth.The existing algorithm ignores the difference of the gradient direction when calculating the gradient amplitude,which will have certain influence on the evaluation result,and there are also shortcomings in the calculation efficiency of color similarity.Based on the above two problems,this paper proposes an improved algorithm for full reference image quality evaluation.Firstly,when calculating the gradient amplitude of the luminance channel,the luminance channel is fused,and the gradient amplitude similarity between the reference image and the image to be measured for the fused luminance channel is respectively calculated,and the final gradient amplitude similarity is comprehensively obtained,and the solution is solved.The problem of existing algorithms with respect to gradients.Secondly,when calculating the color similarity,this paper combines two color channels to calculate the color similarity,which reduces the complexity of independently calculating the color similarity of the two color channels and increases the computational efficiency.This way of calculating color similarity also improves the overall algorithm performance.Finally,the proposed algorithm combines visual saliency,improved gradient magnitude and blended color similarity in four widely used image quality evaluation databases.The experimental results show that the proposed algorithm is more competitive force.In recent years,the non-reference quality assessment algorithm based on convolutional neural network begins to refer to the full reference quality assessment algorithm.This kind of algorithm obtains the distortion map of the image by prediction,but the reference algorithm is based on the basic feature algorithm,and no visual saliency is applied.This algorithm does not fit well with the perception of image quality by the human visual system.In this paper,visual saliency features are introduced on the prediction based on basic features.After extensive experiments on the image quality assessment database,the results verify that the proposed image quality assessment algorithm based on convolutional neural network is more suitable to the human visual system's perception of image quality.