Research On Image Feature Extraction Algorithms Based On GPU

With the rapid development of computer networks and multimedia technologies, there are more and more images in today’s Internet, which are used more and more widely in many fields. How can we realize image retrieval fast and accurately have become the issue that we focus on more and more closely. Through the promotion of the parallelism of algorithms, the efficiency of image retrieval can be improved, which is one of the solutions to this problem. But due to the limit of Central Processing Unit (CPU) by its own material and design concept, it is not suitable for parallel floating-point calculations of large data sets. This is the biggest advantage GPGPUs. Therefore, based on GPGPUs and CPU improved algorithms for image retrieval has become a research hotspot.This thesis firstly introduces some image retrieval technologies, including image color, shape and texture feature extraction methods, similarity measurement algorithms, and etc. And then, it analyzes the development of GPU, GPGPU and the unified compute architecture of CUDA(Compute Unified Device Architecture). After that, it discusses the related technologies used in detail.In this thesis, two kinds of algorithms based on CUDA for processing image feature extraction have been described in the implementation section, which are global color histogram feature extraction algorithm and gray-level co-occurrence matrix feature extraction alogrithm.Global Color Histogram (GCH) of an image is insensitive to shifting and rotation, and needs only simple computation. According to the feature of this algorithm, it is firstly implemented on CPU. Then, according to different versions of CUDA, two improved methods are proposed which are atomic operation and non-atomic operation.The reduction operation of CUDA is used for getting the Color Histogram, which is also optimized.After that, image texture is investigated. It is about how to retrieve image texture characteristics effectively and describe the images. Images should be made gray before being retrieved. The original image is transformed to grey images. Based on the grey images, the grey matrix can be computed, and according to co-occurrence matrices, feature values can be computed. Here, a version is firstly implemented on the CPU, and then, in accordance with the relationship between feature values, the improved version of CUDA is designed and implemented.Comparisons are made between the solutions based on GPGPU and CPU after the algorithms are implemented. From the results of experiments, it is concluded that the effect of acceleration is very good, without affecting to the rate of recognition.