Performance Comparison of GPU, DSP and FPGA implementations of image processing and computer vision algorithms in embedded systems

The objective of this thesis is to compare the suitability of FPGAs, GPUs and DSPs for digital image processing applications. Normalized cross-correlation is used as a benchmark, because this algorithm includes convolution, a common operation in image processing and elsewhere. Normalized cross-correlation is a template matching algorithm that is used to locate predefined objects in a scene image. Because the throughput of DSPs is low for efficient calculation of normalized cross-correlation, the focus is on FPGAs and GPUs. An efficient FPGA implementation of direct normalized cross-correlation is created and compared against a GPU implementation from the OpenCV library. Performance, cost, development time and power consumption are evaluated for the two platforms. The performance of the GPU implementation is slightly better than the FPGA implementation, and less time is spent developing a working solution. However, the power consumption of the GPU is higher. Both solutions are viable, so the most suitable platform will depend on the specific project requirements for image size, throughput, latency, power consumption, cost and development time.