Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots

 
39 
Chapter 5 
5.
 
CNN template calculation schemes with a particular focus 
on the learning/training based approach through Genetic 
Algorithms 
In this chapter the focus lies on the following research question:
 
“
What are the major 
template calculation schemes of relevance for CNN based image processing? How can these 
calculations be performed in a real-
time high performance computing context?”
 
 
5.1
 
Introduction 
Cellular Neural Networks technology provides a very powerful analog computing 
architecture for a variety of array computation and image processing tasks [94]. From a 
theoretical point of view CNN model offers a huge potential for modeling image processing 
filters and operators on a CNN Universal Machine. Each CNN processor matrix used in 
image processing has a feedback template, a feed-forward template and a bias template. 
These three templates can reconfigure the CNN model without any changes in hardware. 
The most challenging issue is to find an optimum set of proper template values for each 
specific application[98]. Figure 5-1, shows this CNN architecture. Overall, there are three 
major ways to calculate the feed-forward and feedback templates:
(a) The Intuitive method:
This first method is needs intuitive thinking of the designer [99]. Depending on the 
designer’s experience in either processing images or dynamics of arrays, we can have a
template. There is no guarantee to find a template for all image processing operators and 
could be very difficult to find a template for complex solution. Experts are familiar with 
template of basic image processing operators and they can combine different templates or 
performing them on CNN individually one by one. For example, we know the template of 
Laplacian of Gaussian for finding edges and also template for smoothing image. If we 

 
40 
combine these two templates on a control template and feedback template respectively, 
results will be an enhanced image with sharpen edges. 

Download 3,22 Mb.