Performance comparison of image matching algorithm using fpga and gpu

As the internet and technology developing very rapidly, the need for fast data processing is becoming more apparent. Even though the software computation capacity of computers are increasing day by day, their performance in large database operations becomes increasingly inadequate. For this reason, Field Programmable Gate Arrays (FPGAs) and Graphics Processing Units (GPUs) with parallel processing capability are frequently used to accelerate the intensive data processing operations. The most effective method for identifying fingerprint images is the template matching. This method is used to match small parts of the image that match the template image or any other data in the entire image. In this thesis, it is attempted to find the existing image within the database by using normalized cross-correlation (NCC) method. Computers that are used today are inadequate for large database operations, even though their computation capacity to process software is high. For this reason, Field Programmable Gate Arrays (FPGA) and Graphics Processing Unit (GPU) with parallel processing capability are frequently used to implement intensive operations. Image database, used in matching, is constructed with 80 different fingerprint images that have 256x256 size, grayscale format and *.bmp extensions. When using the Verilog programming language to program the FPGA, the CUDA programming language was used to program the GPU. In the written codes, image pixel values are read in order and the correlation values between each compared pairs are calculated by storing them in the memory of the used devices. The processing speed of FPGA and GPU implementations are compared in terms of their speed to calculate correlation values. Calculation of correlation values with the FPGA is 23 times faster than GPU implementation. Keywords: Image Matching, Normalized Cross-correlation, FPGA, GPU.