DeNoising

The main component in this package is the CNoiseCleaner class which implements GPU-based denoising algorithm in much the same way as WaveLab's ThreshWave2 function. However instead of using the CPU it employs OpenCL and the GPU to accelerate the algorithm. In order to fully utilize the GPU there is a limit on the size of the images the algorithm can handle. Maximal possible value for the width and height is 1024 pixels. The algorithm has 5 stages:

1. Simultenous forward Haar transform on all of the rows. Each row is processed by a different work-group on the GPU, such that the total number of workgroups is the number of rows (the height of the image). The name of the kernel which is invoked in this stage is FWT_kernel. The way this kernel operates is the reason for the limit (1024 pixels) on the image size. The number of work-items in a work-group must be half the number of pixels in a row for the kernel to work properly. Increasing the size of the image to 2048 and above would require work-groups with at least 1024 work-items and current GPUs don't have such capability yet.
2. Transposing the image such that the columns of the original image become the rows in the transposed one. The name of the kernel which is invoked in this stage is Mat_Transpose_kernel.
3. Simultenous forward Haar transform on all of the rows in the transposed image. This stage is exactly as the first one, but it actually runs on the columns of the original image. This is the reason why the limit of 1024 pixels is true for both the width and the height of the image.
4. The thresholding step which either invokes Mat_HT_Threshold_kernel for hard- thresholding or Mat_ST_Threshold_kernel for soft-thresholding depending on the type of thresholding requested by the user.
5. Simultenous inverse Haar transform on all the rows in the transposed image, which means the columns of the original matrix. The inverse transform works in much the same way as the forward transform but the kernel, which is called IWT_kernel, performs a sort of an expansion rather than a reduction.
6. Transposing the image back into its original form using the same Mat_Transpose_kernel kernel.
7. Simultenous inverse Haar transform on all the rows in the image. Exactly as the fifth stage but it runs on the rows of the image rather than on its columns.

See NoiseCleaner.h for detailed API.

List of files for DeNoising package:

• DeNoising.sln - Visual Studio 2008 solution file.

• DeNoising.suo - Visual Studio user options.

• DeNoising_1\DeNoising_1.vcproj - Visual Studio 2008 project file.

• DeNoising_1\DeNoising_1_main.cpp - A simple test program which tests the functionality of the GPU-based denoising algorithm. It uses OpenCV to read and display the image files. This program also serves as an example for the usage of the CNoiseCleaner class which encapsulates the denoising algorithm.

• HWT_kernels.cl - Contains OpenCL kernels for various stages of the denoising algorithm. This file must be present in the current directory of the application. The kernels are compiled dynamically during runtime, without them the algorithm will not work.

• Makefile - A makefile for compiling the test application in Linux. Serves as an example and can be further extended as needed.

• NoiseCleaner.cpp - Implementation of the CNoiseCleaner class. See comments in the file for further details.

• NoiseCleaner.h - Header for CNoiseCleaner class.

• *.dat - The files that start with signal contain 1D signal in various sizes, and the ones that start with regression contain the corresponding wavelet coefficients. These file are used to test the forward and inverse Haar transform methods in CNoiseCleaner class. This files are not essential and are needed only when CNoiseCleaner::PerformSelfTest method is activated.

• *.jpg - Test images for the test program in DeNoising_1_main.cpp.

• Utils.cpp - Implementations of various auxiliary functions for working with files and OpenCL.

• Utils.h - Header file for various auxiliary functions for working with files and OpenCL.