Research On No Reference Image Quality Assessment Based On Wavelet Transform

With the rapid development of multimedia technology, digital image has become themain source for the people to receive the information, the correctness of the image qualitywill directly affect the understanding and judgment, which is very important in traffic targetdetection, medical imaging and other applications, however, the image will be distortedinevitably by the acquisition, transmission, compression and some other processes, so theestablishment of image quality evaluation standard has the important research value. Amongthem, image quality assessment is divided into subjective and objective evaluation methods,and the objective evaluation method is divided into full reference, reduce reference and noreference evaluation method, this paper mainly research more practical about no referenceimage quality assessment which is not need original undistorted image, the specific contentsare as follows:(1) The thesis deeply discussed the research background and significance of imagequality assessment, and analysis of the current domestic and international research on imagequality assessment results and development, then introduced the full reference, reducereference, no reference image quality assessment and compared the characteristics of theiralgorithms, which gives key analysis on the most research no reference image qualityassessment.(2) Simply introduced the basic theory of the2-D wavelet transform and the humanvisual system, and then analyzed the similarities between characteristics of image wavelettransform and the human vision.(3) Proposed a simple and direct no-reference Gaussian image quality assessment basedon wavelet decomposition. The algorithm take into account the strong correlation among thenatural images’high frequency in the same wavelet scale which would be reduced with theGaussian distortion deepening firstly, and then combined with the peak signal noise ratio,structural similarity and singular value vector for calculating the differences of sub-bandsrespectively which were the three objective assessment index. The simulation results showthat the proposed method has good consist with the subjective assessment on the threecommon image databases，in addition，compared their pros and cons.(4) In view of the current no-reference image quality evaluation method need to learnedthe type of image distortion in advance, combined with the advantages of BP neural network,this paper put forward a no-reference image quality evaluation method based on natural scenestatistics and BP neural network which was suitable for multiple types. First of all, accordingto the NSS that the undistorted natural image wavelet sub-bands coefficients has the similar linear rule in logarithmic domain. As the distortion usually manifest in the high frequency andlow frequency change is not obvious, used low frequency information combined with thelinear rule to predict the high frequency information, and then forecast information and theactual difference as image characteristics. The algorithm combined the differences and theirsubjective score to training the BP neural network which is for catching the last objectiveassessment model.