Research And Implementation Of Key Algorithms In Image Information Extraction Based On Heterogeneous Computing Systems

In recent years, the image of Information Science has become more widely used than ever, image information extraction technology has also become the research hotspot in the field. Image wavelet transform algorithm and K nearest neighbor classification algorithm, often used as the image feature extraction algorithm and classification algorithm, widely exists in all kinds of image information extraction system and application such as scene recognition, based on the content of the image search engine, criminal investigation of image classification, computer vision, etc. However with the practical engineering image of the quantity and scale of growing, and people’s demand on high performance and low power consumption calculation is becoming more and more urgent, the existing implementations of image wavelet transform algorithm and K nearest neighbor classification algorithm are increasingly stretched. Using a new FPGA architecture for heterogeneous computing system for exploring provides the possibility of efficiently implementing the algorithm. Compared with the super computer or distributed computing system solutions, using FPGA architecture for heterogeneous computing system can achieve low cost, less land occupation and better repair advantage. Furtherly, implementation of heterogeneous system using FPGA has the advantage of low power consumption and high energy efficiency compared with GPU. In summary, this paper presents a FPGA architecture based model of heterogeneous system in the OpenCL parallel accelerating research, and implement wavelet transform of image and K nearest neighbor classification, which are two key algorithms of image information extraction technology.In this paper, two strategies of image wavelet transform algorithms are used on heterogeneous computing system with FPGA. One implementation is to use a row convolution kernel and column convolution kernel, and the other is to reuse the convolution kernel and matrix transpose kernel.Based on the combination of parallel odd even sort and parallel bubble sort, this paper presents a k nearest neighbor classification algorithm for heterogeneous computing system with FPGA. The algorithm use two kernels, similarity computing kernel and similarity sorting kernel. In similarity sorting kernel the K parallel odd even sorting and K parallel bubble sort are used in different stages, and with the reuse of the two kernels the implementation is accomplished.Finally, based on FPGA+CPU heterogeneous computing system, the image wavelet transform algorithm and K nearest neighbor classification algorithm are implemented to accelerate the computing task. With optimization method provided by AOCL tools, the vector processing and multi pipeline replication optimization are used to improve the system performance and utilization rate of resources. In contrast to using GPU+CPU platform, we made a detailed comparative analysis from the calculation speed and the energy conversion efficiency of the algorithm. The test using a public gallery and data sets, the results are consistent with theoretical expectations. In this paper, the design of the image wavelet transform heterogeneous parallel acceleration scheme with FPGA platform compared to the GPU platform improved energy efficiency ratio by 23.7%. And with the combination use of parallel odd-even sort and parallel bubble sort, the energy efficiency ratio of heterogeneous implementation of K nearest neighbor classification algorithm increased by 18% compared to the parallel bubble scheme.