No Reference Image Ambiguity Estimation,

Visual information is the main means of human access to information, it is obtained through human visual perception system. In visual information, image information is the most important part. Along with development of personal computers, digital communications, multimedia and network technology, digital images and digital videos is increasingly becoming one of the most important carriers of information, has penetrated into people's daily lives, reach millions of households. In the digital image acquisition, processing, coding, storage, transmission and reconstruction of each step, image quality are impacted, how to assess the image quality has become a fundamental and challenging problem in image processing and computer vision. As a most important part of No Reference Image Quality Assessment is an important part, No-Reference Image Blur Estimation is the topic in this thesis. Based on the studies of image distortion characteristics, image blur process modeling, edge extraction, HVS systems and saliency features, this thesis also proposed a template matching based No-Reference Images Blur Estimation method and visual salience-weighted based method Image Blur Estimation.This thesis will focus on analysis of the edge type in the images, Gaussian template matching method and visual saliency characteristics. The main work and innovation are summarized as follows:1) Apply template matching method to the image blur estimation. Proposed a template matching based image blur estimation method. Designed adaptive connecting method for fractured edges. Designed a Gaussian template set, use each template in this set to matching with the edge gradient curves for getting edge gradient curve at peak signal to noise ratio. Draw probability distribution curve of gradient curve standard deviations (STDEV), by the position of peak value of PDF calculate the image point spread function standard deviation. The theory of template matching is matching filtering which has the advantage of detecting signal under best signal to noise ratio. Suppose templates in Gaussian template set as signal to detect and edge gradient curves as filter, the output of filter will be best matched template and the STDEV of this template is the STDEV of the edge gradient deviation.2) Apply Image saliency theory to image blur estimation, and optimized the image blur estimation algorithm. It makes estimated blur degree correlate well with subjective image blur marks. Image saliency is a mechanism of HVS, that when HVS viewing image, some interesting area will attract more attention. In general, these areas are more structured. Based on the multi-scale image saliency theory this thesis proposed an optimized image blur estimation algorithm.