Video Denoising Based On Surfacelet Transform Using Statistical Models

Along with the development of the information age and the popularity of video processing, video quality has been increasingly and highly demanded. Video is often corrupted by noise during collection, acquisition and transmission. The noise is the main factor that influencs video quality and seriously affects to information extraction of the images. So, it is required to remove the noise from video before analyzing the video. The major works can be summarized as follows:(1) In this paper, a video denoising method using coefficient shrinkage and threshold adjustment based on surfacelet transform (CSTA-ST) is proposed, which processes multiple frame video as an ensemble. The spatially estimated energy (SEE) based on weighted model, is constructed using the spatial correlation. Each ST coefficient has a corresponding SEE value, and ST coefficients are grouped according to SEE value. According to the similarity of ST coefficients in the same group, the threshold of each ST coefficient is determined. In addition, according to the neighborhood information of ST coefficients, the coefficient shrinkage parameter (CSP) and threshold adjustment factor (TAF) are determined and the ST coefficients are modified. Finally, the denoised video is obtained by inverse surfacelet transform using the coefficients. In experiments, video sequences with noise are tested, and the denoised results of the proposed method are compared with that of the relevant denoising methods.(2) In the forth chapter, a video denoising method using Gaussian scale mixture model(GMM) in the surfacelet transform domain(ST). First, a Bayesian threshold is introduced in order to classify the ST coefficients. Then we use a exponential distribution model and Gaussian scale mixture to capture the correlation of the classified ST coefficients. Finally, the coefficients are shrinked based on the model. In experiments, video sequences with noise are tested, and the denoised results of the proposed method are compared with that of the relevant denoising methods.(3) This paper also proposes a new video denoising method based on the NonsubSampled Surfacelet transform domain and bivariate model under the framework of Bayesian MAP estimation theory. In this algorithm(BM-NSST), we use the NonsubSampled Surfacelet transform domain advantages of translation-invariant and multidirection-selectivity to exploit the intra-scale and inter-scale correlations of NonsubSampled Surfacelet transform domain, and elaborates the method of noise estimation. Finally, the denoised video is obtained by the inverse surfacelet transform using the shrunk coefficients. In experiments, video sequences with noise are tested, and the denoised results of the proposed method are compared with that of the relevant denoising methods.