Figure
5-9: Removing noise from the image
54
An image having a tube is fed to the CNN, as shown in Figure 5-10(a). The aim is to fill the
tube with a given dotted pixel. Figure 5-10(c) shows the output as specified by target.
(5-8)
Figure
5-10: Filling a long tube with a dotted pixel
In Figure 5-11, a changing gradient image was provided as input. The aim is thresholding
up to a specified limit as shown in the target image. Figure 5-11(c) shows the output image
for the following set of parameters.
(5-9)
55
Figure
5-11: Thresholding to a specified limit
Obstacle detection was performed by a sequence of stereo images using Cellular Neural
Networks. The CNN parameters were determined by a genetic algorithm based on real
number chromosomes. Using real number chromosomes, repeated encoding and decoding
of the chromosomes is not required. Unlike binary chromosomes, relatively smaller
chromosomes are produced. This approach was successfully implemented for obstacle
detection and also it was found to produce good results for many other image processing
tasks. In future, we aim to implement this approach on hardware level for enhancing
efficiency and speed.
5.4
Real-time computing issues for the genetic algorithm based
CNN template’s calculations
Adaptive image processing and image analysis is very important for ADAS concepts.
Processing image under different and environmental visual conditions such as fog, rain,
and sun in background is not trivial. To overcome problems in this form we need to involve
adaptive image processing techniques. Our CNN platform can process images with a
performance of 100 FPS. For dynamic processing, CNN needs dynamic templates or set of
templates. To have a fast response, CNN needs to access different templates very fast.
There are three ways for real-time template accessing. Template pre-calculation is one
solution. There are many standard and classic templates like thresholding, contrast
enhancement, dilation, opening, erosion, closing, find edging, median filter, etc that can be
pre-calculated and stored in a list. And depending on the situation the procedure system
56
can call different templates. Figure 5-12 shows this structure for pre-calculation of
templates.
|
