Research On Image Quality Assessment Method Based On Visual Perception And Natural Scene Statistics

Image quality assessment technology can effectively evaluate received images,and avoid the loss caused by image distortion.In this paper,we propose three objective image quality assessment methods based on the human visual perception and statistical features of natural scenes.(1)Based on visual perception,a FR-IQA method is proposed to evaluate the quality of single-distortion images.According to the visual perception of the human eye,which is extremely sensitive to the edge of the image,the image is segmented by perceptual importance,then filtered by the directional contrast sensitivity function.Finally,the weighted SNR of the reference image and the distorted image is calculated to construct an objective image quality index.The experimental results show that the proposed method has a high degree of subjective consistency for the evaluation of distorted images.(2)Based on visual mutual information,a FR-IQA method is proposed to evaluate the quality of multi-distortion images.First,the reference image and the distorted image are visually pretreated through the steerable pyramid decomposition and contrast sensitivity function to simulating the multi-channel characteristics and contrast sensitivity characteristics of the human eyes.Next,a Gaussian scale mixture model is constructed for the reference image,and an image distortion model including signal attenuation and additive noise is constructed for the distorted image.Then,a visual distortion model is constructed for both the reference image and the distorted image.Finally,the mutual information between the reference image and the distorted image after the above model passed is calculated to construct the full reference quality assessment index of the multi-distortion images.Experimental results show that the proposed method is superior to the existing methods in evaluating the quality of multi-distortion images.(3)Based on AdaBoost_BP neural network,a NR-IQA method is proposed in wavelet domain.The generalized Gaussian model is fitted to the subband of the distorted image in the wavelet domain,and the model parameters are extracted as features of the natural scene statistics;mean and skewness of local entropy of subbands coefficients are extracted as information entropy features.Combining these two kinds of features to construct a 36-dimensional image feature vector.Two AdaBoost_BP neural networks are trained as the image distortion classifier and the image quality grader using the known distortion image feature matrix,the distortion type and the subjective scores.Through training and learning,the quality score and distortion type judgment of the no reference image are achieved.The experimental results show that the proposed method has high accuracy for the distortion type judgment and the quality score has good subjective consistency for all types of distortion.