Research On Image Motion Deblurring And Wavelet Threshold Denoising Algorithm

Image restoration and denoising are essential topics in image processing, which can greatly improve image quality. This dissertation primarily discusses the technologies of the motion blurred image restoration and the image denoising via wavelet transform. The main work of this dissertation can be summarized as follows:1. Taking into account the traditional image quality evaluation method can not reflect the visual perception effectively, we proposed a improved SSIM method on the basis of orientation information measure. The experiment results illustrate that the proposed metric has good correlation to the subjective perception and can reflect the image quality effectively.2. For restoration of motion blurred image, current algorithms mainly focus on a shift-invariant linear blur, which are not applicable for spatially variant motions. A novel method for rotational blur identification and restoration is proposed in this thesis. Firstly, efficient arcs are extracted from binary edge image and rotational center is detected using the LSM (Least Square Method) circle fitting. Secondly, blurred image is transformed to compact polar coordinate and kurtosis-parameter curve is established to detect rotational angle. Thirdly, an improved constrained least square method is applied to recover the degraded image. Experiment results show that our method is effective.3. After studying the theory of wavelet image denoising, we addresses the research work on threshold denoising algorithm and present three denoising methods. First, a new threshold function based on wavelet threshold denoising is presented. The method can overcome the shortcoming of the hard threshold with discontinuous and solve the problem of the bias in soft threshold method. Second, we propose a improved NeighShrink method using generalized cross validation (GCV) theorem to estimate optimal threshold and neighbouring window size. Experiments demonstrate this algorithm outperform NeighShrink and can get better denoising results than the methods available. At last, since wavelet transform can not capture geometric regularities of real images, we introduce multiscale geometric analysis to reduce noise. In particular, by introducing the improved NeighShrink into the Contourlet domain, we develop a new image denoising method. This algorithm can not only remove noise, but can effectively maintain the edges of images. It outperforms other denoising methods in terms of PSNR and visual qualities.