INTERNATIONAL JOURNAL OF TECHNOLOGY ENHANCEMENTS AND EMERGING ENGINEERING RESEARCH, VOL 2, ISSUE 7 1 ISSN 2347-4289 Copyright © 2014 IJTEEE. Frame By Frame Digital Video Denoising Using Multiplicative Noise Model M. Morshed, M. M. Nabi, N. B. Monzur Department of Electrical and Electronic Engineering, Ahsanullah University of Science and Technology, 141-142, Love Road, Tejgaon I/A, Dhaka 1208, Bangladesh. Email: monjurm@aust.edu ABSTRACT:The sparse representations of images have achieved outstanding demising results in recent days. But noise reduction in digital videos remains a challenging problem. In this communication we considered the coherent nature of the video frames for image processing. The imaging model shows that the video frames are corrupted by multiplicative noise. Simulation results carried out on artificially corrupted videos' frames and demonstrated performances of five previously available filtering approaches. Keywords : Digital Video Denoising; Multiplicative Noise Models; Peak Signal to Noise Ratio; Synthetic Aperture Radar; Structural Similarity Index etc. 1 INTRODUCTION VIDEO signals are considered as a sequence of two- dimensional images, projected from a dynamic three- dimensional scene onto the image plane of a video camera. Luminance and chrominance are two attributes that describe the color sensation in a video sequence of a human being. Luminance refers to the perceived brightness of the light, while chrominance corresponds to the color tone of the light. Nu- merous still images and video denoising algorithms have been developed to enhance the quality of the signals over the last few decades [1]. Many of the algorithms are based on proba- bility theory, statistics, partial differential equations, linear and nonlinear filtering, spectral and multiresolution analysis. How- ever, image denoising can be extended to a video by applying it to each video frame independently. Depending on various signal-processing problems various algorithms have been proposed mainly for image denoising [2]. A human observer cannot resolve fine details within any image due to the pres- ence of speckling. The available techniques are mostly based on noise suppression techniques in the post-image formation type; use computer simulation to suppress the signal as well as its speckles. The property of image sparsity is an important key to denoise image and video signals as well. Sparsity also resides in videos. Most videos are temporally consistent; a new frame can be well predicted from previous frames. The idea of combining multiple images to get a desired one is called image fusion and can be used to produce a denoised video. Video signals are often corrupted by additive noise or motion blur. Often, the noise can be modeled effectively as a Gaussian random process independent of the signal. Although the state of the art video denoising algorithms often satisfy the temporal coherence criterion in removing additive white Gaus- sian noise (AWGN) [3]-[6]. Normally all coherent imaging processes, such as, synthetic aperture radar (SAR) and nar- row-band ultrasound suffer from speckle noise. The SAR im- ages are available in two formats. One is amplitude format and the other one is in intensity format. The magnitude of the speckles follows the Rayleigh probability distribution and cor- responding phases follow uniform distribution [7]. The speckle intensity is described by a negative exponential distribution. By multilook averaging the undesirable effects of speckle in a SAR image can be easily reduced. For the case of intensity data the statistical distribution of the resultant speckle in the degraded resolution image is given by the Gamma distribution and in case of amplitude data multiconvolution of the Rayleigh probability density is considered. Speckle in SAR images is generally modeled as multiplicative random noise [8], whereas most available filtering algorithms were developed for additive white Gaussian noise (AWGN) in the context of image denois- ing and restoration, as additive noise is most common in imag- ing and sensing systems. Fig. 1. General video denoising framework. Video denoising is normally done with some linear or non-linear operation on a set of neighboring pixels and the correlation between those pixels available in spatio-temporal sense. The best video denoising can be achieved by exploiting information from both future and past frames. But this leads to an additional delay of at least one frame which is undesirable in some real- time applications. For this reason, many algorithms exploit information from usually the current frame and one or two previous frames. In image frame denoising algorithms focus to find the best compromise between noise removal and preservation of important denoised image frames. Here each frame is independently processed. That is why for optimal filter performance the spatio-temporal properties of the processed noisy image frames are taken into consideration. The general framework for video denoising is illustrated in Fig. 1. An accurate modeling of noise is necessary in order to estimate noise-free spatio-temporal sequence structures. To distinguish between the noise and the noise-free spatio-temporal correlations in the image frames the information concerning the noise and the noisy input frames are combined together. In this way, the spatio-temporal structures can be estimated in a noise-insensitive manner and consequently enable an efficient noise removal with the preservation of all the important spatio- temporal sequence features [9]. In this communication, we considered multiplicative noise models for video signals