Color Image Quality Assessment Based On Color Perception Characteristics

In the field of computer vision,we always simulate the human eyes to identify and track the target based on the gray images.However,we live in a colorful world,most of the image information received by the human eyes is color.Therefore,the computer vision algorithm based on color images analysis can analyze the image content scene more accurately,meanwhile improving the accuracy of image recognition.But the quality of the color image is the basic condition for determining the performance of image recognition and tracking algorithms.Therefore,it is of great significance for the computer vision algorithm to design color image quality assessment.Color image quality assessment is using the mathematical models to predict the image quality automatically.Most of the traditional quality assessment algorithms are based on the image structure information to design models,ignoring the image degradation caused by the distortion or loss of color information,which leads to poor consistency of this kind of algorithms with human subjective perception.At the same time,due to the diversity of color changing and the complexity of color space,the research of color image quality assessment is still in development stage.We have carried out in-depth research on this issue about color image quality assessment based on the characteristics of color frequency domain,the redefinition of color space and the color visual characteristics.The main research work and contributions in this thesis are as follows:(1)A color image quality assessment using statistics of color descriptors in the DCT domain is proposed.The algorithm uses the distribution of the color descriptor DCT coefficients to characterize the degree of the color image degradation,and a novel color image quality assessment is proposed by combining the statistical model of natural scenes with the DCT transform.First,the color descriptors of the images are extracted,including saturation,hue,and luminance.Meanwhile the discrete cosine transform is used to extract the frequency features,analyzing the feature changes with the color image quality under different distortions.Then the mapping function of the image features and the quality are studied by the support vector regression.Experimental results show that this method achieves ideal consistency with human subjective quality score.(2)A color image quality assessment based on quaternion DCT subband similarity is proposed.The method uses three imaginary of a quaternion to represent the three-channel value of the color image and describes the color image as a whole.First,a quaternion matrix is used to express a color image,and a quaternion discrete cosine transform is performed on it to obtain the quaternion spectral coefficients of the image,which are then divided into different subbands according to different spatial frequency positions.Finally,the subband similarity between the reference image and the distorted image is weighted and summed to obtain the image quality.The experiment results of the standard databases show that the proposed method has more advantages in terms of consistency and effectiveness.(3)A color image quality assessment based on color perception is proposed.The accurate quality evaluation of color images should imitate the color perception of human visual system,and abstract the color perception characteristics into color features to construct the algorithm model.This method combines feature fusion and classification algorithm to transform the quality prediction problem into preference learning problem,and then designs color image quality method.First,we extract the color harmony,color contrast,colorfulness and the related luminance features for the preference image pair.Then,the preference learning problem is solved is based on the machine learning classification algorithm.Finally,the prediction of color image quality is performed based on the voting strategy.This method has achieved higher subjective and objective consistency.