A fast method for video deblurring based on a combination of gradient methods and denoising algorithms in Matlab and C environments

Abstract

In this paper the degraded video with blur and noise is enhanced by using an algorithm based on an iterative procedure. In this algorithm at first we estimate the clean data and blur function using Newton optimization method and then the estimation procedure is improved using appropriate denoising methods. These noise reduction techniques are based on local statistics of clean data and blur function. For estimated blur function we use LPA-ICI (local polynomial approximation - intersection of confidence intervals) method that use an anisotropic window around each point and obtain the enhanced data employing Wiener filter in this local window. Similarly, to improvement the quality of estimated clean video, at first we transform the data to wavelet transform domain and then improve our estimation using maximum a posterior (MAP) estimator and local Laplace prior. This procedure (initial estimation and improvement of estimation by denoising) is iterated and finally the clean video is obtained. The implementation of this algorithm is slow in MATLAB1 environment and so it is not suitable online applications. However, MATLAB has the capability of running functions written in C. The files which hold the source these functions are called MEX-Files. The MEX functions allow system-specific APIs to be called to extend MATLAB's abilities. So, in this paper to speed up our algorithm, the written code in MATLAB is sectioned and the elapsed time each section is measured and slow sections (that use 60% of complete running time) are selected. Then these slow sections are translated to C++ and linked to MATLAB. In fact, the high loads of information in images and processed data in the for loops of relevant code, makes MATLAB an unsuitable candidate writing such programs. The written code our video deblurring algorithm in MATLAB contains eight for loops. These eighth for utilize 60% of the total execution time of the entire program and so the runtime should be potentially decreased by considering these loops in C. However, after implementing eighth for in C using the MEX library, the measured run time unexpectedly increased. According to the timing in MEX, this runtime increscent is not because of connecting MATLAB to MEX but it is related to loops. Our simulation shows implementation of a loop in C++ takes two times more than the same loop in MATLAB. In spite of C functions, the OpenCV2 functions take less time, so OpenCV that is an open source computer vision library written in C and C++ (and it is an active development on interfaces MATLAB) would be useful getting more speed. After implementation the for loops of our algorithm using OpenCV library, our simulations show the run time decreases a lot.