Study On Image Fusion And Denoising Based On Shearlet Transform

Multi-sensor Image fusion (including high spatial resolution Panchromatic image and multispectral image fusion and multi-focus image fusion, etc.) and image de-noising (including Gaussian noise removing and random-valued impulse noise removing, etc.) are the current research focus in the field of image processing. Its difficult for these existing transform domain high spatial resolution panchromatic image and multispectral image fusion algorithms to achieve a good balance between improving spatial resolution and maintaining spectral information; these existing transform domain multi-focus image fusion algorithms cannot select all the transform coefficients correctly, leading to the sharpness of the fused image decreases to some degree compared with the focused regions in the source images; these existing Gaussian noise removing algorithms may blur the edges and details of the image while removing Gaussian noise from it; these existing random-valued impulse noise removal methods cannot recover all the geometric features of the data that were corrupted by noise, resulting in jagged edges and other types of distortion appeared in the de-noised image.To address these existing problems in the field of image fusion and denoising, the author took the translation invariant Shearlet transform as the main line, conducted in-depth research in its application in multi-sensor image fusion and image noise elimination in this dissertation. The main research work and innovation of this dissertation can be summarized as follows:(1) For the problem of high spatial resolution panchromatic image and multispectral image fusion algorithms cannot achieve a good balance between improving spatial resolution and maintaining spectral information, a novel fusion algorithm based on segmentation and IHS color space domain Translation Invariant Shearlet Transform is proposed. The algorithm combines IHS transform and TIST together, using Mean-shift image segmentation method for high spatial resolution panchromatic image segmentation; Based on the variance of each region, classify these regions into two categories:regions need to enhance details and regions need to maintain the spectral characteristics of the multispectral image, from which to guide the fusion of TIST domain bandpass directional subband coefficients. For TIST domain low frequency sub-band coefficients, the fusion strategy based on local fourth-order correlation coefficients is utilized; for bandpass directional subband coefficients, the fusion strategy of injecting the coefficients of high spatial resolution panchromatic image is used in the regions of boundary and the regions that need to enhance details; in the regions that need to maintain the spectral characteristics, the fusion strategy of keeping the coefficients of the multispectral image is utilized. This fusion algorithm can achieve a good balance between improving the spatial resolution and maintaining the spectral characteristics of the multispectral image, make the fused multi-spectral image not only keep the spectral information of the source multispectral image well, but also improve its spatial resolution effectively.(2) For the drawbacks of those multi-scale decomposition based multi-focus image fusion methods (First, commonly used multi-scale decomposition and reconstruction methods all have different levels of error, the existence of such errors may lead to the loss of parts of the source image’s useful information; Second, the complex content of images in the multiscale transform domain may lead to the fused transform coefficients are selected by mistakes, causing the reduction of the sharpness of the fused image to some extent with respect to the focused regions of the source images.), a TIST domain multi-focus image fusion algorithm based on the detection of the focused area of the source multi-focus images is proposed. Firstly, by means of a simple TIST domain multi-focus image fusion algorithm, an initial fused image is obtained; and then, based on the local similarity between the pixels of those source images and the initial fused image, the focused area of each source image is detected; and accordingly, all of the source image pixels are divided into three categories:the interior of focused region, the transition region and the interior of defocused region, whereby to guide the final fusion of TIST domain sub-band coefficients. This algorithm can integrate more useful information in the source images into the fused image, avoiding the introduction of "artifacts" and other false information, improving the quality of the fused image effectively, and thereby obtain a fused image which is more similar with the standard reference image.(3) For the removal of Gaussian noise from images, a TIST domain Gaussian Scale Mixture (GSM) Model based image denoising method is proposed. According to the distribution of the Shearlet coefficients of images have non-Gaussian and heavy-tailed characteristics, this method use the GSM model to describe the local distribution of the TIST coefficients of images, then estimate the noise-free TIST coefficient located in the center of the neighborhood by means of the Bayes Least Squares (BLS) estimation method. Experimental results show that the proposed method can suppress the Gaussian white noise and keep the edges and texture information effectively, thus improve the objective and subjective effects of the denoised image.(4) For the detection and removal of random valued impulse noise from images, a powerful refinement of the Highly Effective Impulse Noise Detection algorithm (HEIND) is proposed first by introducing an efficient iterative technique; and on this basis, a powerful two-stage random valued impulse noise removal method which based on the improved HEIND method and the Shearlet based image inpainting method is proposed. This method takes these noisy pixel locations which are detected by the improved HEIND method as the area that need to repair, then use the Shearlet based inpainting method to restore the missing information at those pixels that have been identified during the first stage, thereby effectively restore the geometric features that were destroyed by random valued impulse noise while removing them, thus better protecting the edges and details, make the denoised image has a better visual effect.