82
implement these operations only by cascading convolution operators and some Boolean
arithmetic operations. For the complex algorithms like “skeleton”, CNN has clearly a very
good potential and it is easy to find a template to extract the image skeleton; see Refs [18,
129] [93, 96, 130].
7.11
Conclusion
In this chapter, a new method for CNN emulation on FPGA for real time machine vision
applications has been proposed. The system is implemented on Xilinx XtremeDSP kit 3400,
which is very flexible for video and image processing applications. The video is tested on
DVI and Camera in connection with 1024×768 pixels and 60 fps monochrome for direct
convolution and near 24 fps CNN emulation. The maximum frequency of the system is
200MHZ. The Computation of pixel operations for convolution method does only take two
clock pulses. And for the CNN implementation the number of clock pulses needed for a give
processing did depend on the accuracy needed to repeat the DDA iterations. The whole
design has been made by ISE 11.2, EDK 11.2 and Impulse CoDeveloper 6.3. The main
features of the convolution technique is that we don’t need to access to the ext
ernal
memory and the frame rate is high. That’s why it takes only two clock pulses. For
implementing complex filters such as nonlinear image processing, CNN-based processing is
clearly much better than convolution operations. To achieve the same result by cascading
some convolution operators would result is a significantly slower process than using only
one CNN operation processing.
