A Graphical Processing Unit Based on Real Time System

Abstract

The field of real-time image processing focuses on analysing and enhancing images in real time. An effective way to accomplish this is to utilise both the central processing unit (CPU) and the graphics processing unit (GPU) at the same time, to achieve the highest possible performance. The CPU manages tasks such as sequential operations and input-output, while the GPU is used for operations that can be done in parallel. The real-time processing of medical digital images requires the processing of an image with a response guaranteed within a specified time. This time period should be within a couple of seconds or less, so that images can be analysed and manipulated in a real-time medical scenario. Real-time image processing has gained significance due to its widespread use in communication schemes involving video conferencing, video calls, innovative media, and digital and mobile cameras. Because of this, image identification has recently become an important research topic. The goal of this work is to develop a real-time digital image processing system aimed at feature identification and classification. This thesis presents a method which seeks to reveal normal and malignant tissues. A method of digital image processing via real-time processing is designed to allow radiologists to detect abnormal tissues or cells, and then determine whether a tumor is a carcinoma or benign. Although radiologist decisions remain the preferred way of detecting non-palpable cancer, the method in this thesis aims to make detection by radiology trainers and radiologists easier and more accurate than is the case with traditional techniques. The research described in this thesis involves a toolbox with a number of different filters, including classical high- and low-pass filters, as well as various novel morphological filtering tools. Edge and feature detection algorithms have also been added. This thesis presents the first phase of the removal of noise from noisy images (having a signal-to-noise ratio of 10% or more). Simulated noise of different types is added to original images and then a variety of filters including mean, median, erosion, dilation, opening and closing filters are applied. These filters are used to denoise the original images. Erosion and dilation filters are the two basic filters in the area of mathematical morphology. Although they are usually used at the binary level, in the present research they are used at the grey-scale level. Mean and median filters function in a similar manner, except that median filters preserve more important details in a processed image than is the case with mean filters. In addition to the above-mentioned filters, Laplacian filters are also utilised to increase the contrast of edges, and thresholding techniques are then applied for a first attempt at feature identification. Although the initial work was done in MATLAB, the algorithm developed here is also implemented using CUDA on graphics processing units, with the goal of implementing the system in real time or near real time. Moreover, some algorithms related to segmentation and automatic identification features have been developed in CUDA.