Research On Stereoscopic Image Blind Quality Assessment Method In Application

With the continuous development of computer technology,people require higher and higher things towards the image quality.Therefore,effective quality assessment of 2D images and stereo images has become the current research work.Referring to the research results of computer vision,this paper proposes three image quality assessment models based on machine learning and sparse representation.The main contents are as follows:1)The image blind quality assessment based on recommendation application is proposed for self-images.Firstly,this paper establishes a selfie database and subjective test.For this database,this paper uses an unsupervised quality evaluation model to evaluate the quality of naturalness,then detects the face based on the skin color detection algorithm,and the brightness mean of all the pixels in the face detection area to evaluate the local brightness,then locates the face centroid according to the face detection frame,and then uses aesthetic criteria to evaluate the layout quality.The decision-making mechanism calculates the overall quality of selfie.Finally,the performance of the proposed algorithm is evaluated by the analysis of receiver operating characteristic curve.Results show that the proposed algorithm is highly consistent with the subjective scores of human eye.2)The hybrid image blind quality assessment is proposed for hybrid stereo images.For further research,this paper establishes a hybrid stereo image database.For this database,this paper divides the distorted image into matching area and mismatching area according to the binocular characteristics;Then,the matching area of left and right viewpoints is fused into a monocular image according to the binocular competition characteristics;Then the mismatching area and the monocular image are decomposed by wavelet transform;Then the natural statistical features are extracted by multiple decomposition coefficients;Finally,mapping features into image quality by using support vector regression algorithm;In testing stage,stereo image quality is predicted according to the established regression model.Results show that the proposed algorithm is better than the existing methods and highly consistent with human eye subjective score.3)A stereo image evaluation method based on union dictionary is proposed for multi-distortion stereo images.In this paper,four kinds of distortions are trained separately.Firstly,according to the principle of binocular competition,the matching regions of left and right viewpoints are fused into a monocular image.In the training stage,the left distorted image and monocular image is normalized and divided into 8×8 non-overlapping blocks.Then,the original left viewpoint and the distorted left viewpoint,the original monocular image and the distorted monocular image are separated into 8×8 non-overlapping blocks by using FSIM respectively.Then multi-modal dictionary is learned from feature image blocks and FSIM mapping blocks.Finally,feature dictionary and quality dictionary are trained.In the testing stage,orthogonal matching pursuit algorithm is used to calculate the sparse coefficient,and then according to the different combination of sparse coefficient and quality dictionary,the quality matrices of four kinds of distorted images in left mismatched region,right mismatched region and cyclopean are obtained respectively.Then,the corresponding quality values are obtained using mean fusion algorithms and the quality values of stereo images are obtained using the weights.This method achieves good results on asymmetric and symmetrical hybrid stereo image database and several classical databases,and keeps a high degree of consistency with human eyes.