Research On Blurred Image Restoration And Image Quality Assessment Algorithm

Blurring can reduce image details and make image edge smooth, which has serious impact on the military target detection, medical lesions identification and vehicle license plate recognition, etc. To solve these problems, the mechanism of image blurring, image restoration algorithms and image quality assessment methods are researched as follow.Firstly, a whole image restoration framework is established for real-world deblurring. A blur identifying process is proposed to decide which image needs to be restored. Also, a image quality assessment step is performed before and after image restoration, which is used to ensure the image restoration system having a reliable output.Secondly, the spatial and frequency features which can distinguish between sharp and blurry images are studied to form blur identification algorithms. In the spatial domain, the heavy-tailed distribution of nature image is used to describe the sharp image feature, the blur identification is realized by analyzing the histogram of image gradients. In the frequency domain, the image discrete cosine transform coefficient distribution is analyzed to come up with a new index to indentify blurry images. Also, the two kinds of identification algorithms were compared by simulation experiments.Thirdly, the linear motion blurring mechanism is studied to find image restoration algorithm of the best performance. The blurring parameter estimation algorithm of spatial, frequency and cepstrum domain are analyzed. A novel cepstrum extremum detection algorithm is proposed to make up the defects of line detection algorithm based on Hough transformation. At the same time, the ringing suppression method is studied to improve the performance of Richard-Lucy algorithm.Furthermore, blind image restoration algorithms are studied to restore defocus blurring and complex motion blurring. For existing algorithm, insignificant edges make blur kernel estimation process uncertainly, which make the restored image unclear. In order to solve this problem, a novel normalized sparse regularizaiton algorithm based on saliency gradients selection is proposed. Simulation experiments are designed on synthetic and nature blurring images to verify the performance of the new algorithm.Finally, the image quality assessment algorithms are researched and an improved structure similarity metric is proposed. In order to obtain a no-reference metric, a re-blur approch is adopted to the improved full-reference index. Considering the ringing artifacts in the restored image, a co-occurrence pixel detection algorithm is performed to measure this degradation and genarate a quality degradation factor. At last, a novel no-reference image quality assessment algorithm with luminance, contrast, structure, sharpness and ringing metric is designed to evaluate the restoration result.