71
industrial systems [26, 49, 50]. In the classical video processing platforms we can use DSP
or CPU core for manipulating/processing pixels. For process one frame of data, the system
generally has to fetch both the data and the program to either CPU or DSP,
perform
required mathematical operations and then store the result(s) back into the memory. The
system must handle the high priority interrupts at the same time. And all these extra cycles
will add to the total number of cycles involved in the processing each pixel of image [121].
The main weakness of these traditional systems is clearly the low speed related to the high
processing time. Due to the sequential architecture and the programs, the system cannot
manipulate pixels in a real pipeline model. Therefore, we must design a suitable
architecture with a pipelining potential. FPGA is one the best
candidates for pipelining
video processing. With newest FPGA technologies it is possible to design a multi-functional
and high performance video processing system. New FPGA technologies have made them
much faster and denser than before. XILINX Vertex technology provides a large two-
dimensional array of logic and programmable block sets, which
contain lot of dedicated
memories and flip-flops. Having such facilities and infrastructures, one can easily map the
image on this grid for further image processing[49]. This implementation presents a real-
time video processing platform involving two concepts: a) direct convolution based image
processing, and (b) CNN based image processing. Designing a proper image processing
platform is extremely significant. Thus, we have to design a robust and flexible
architecture. The main parts of a real-time platform
are knowingly capturing video,
buffering video streams, video stream processing and finally video output controller. All
these parts are considered in the platform design of this implementation.
