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

 
31 
Figure 4-1 shows a simple feed-forward neural network with four inputs and four cells as a 
hidden layer and one output. Information can flows from input layer to hidden layer and 
then to output layer through the connection which are generally called links or synapses. 
Each layer except the input layer has an activation function. Flow of information from 
input to output called feedforward. After training, the operation mode is feedforward 
mode. Means loading information in input layer and system calculate the proper output 
according to weight and topology of the network [75]. The most common learning method 
for teaching and tr
aining the MLP networks is backpropagation. It’s a supervised learing
method and it is derivation of delta rule method. For training we need to perform 
propagation phase and then updating weight. In propagation phase, we need to perform 
forward propagation of training pattern and get the result from output layer activation 
functions. After that, performing backpropagation of these output data by considering 
target in order to generate the delta of output and hidden neurons. For updating the 
weights we need to multiply output delts of each layer and input activation to calculate the 
gradient of the weight and then updating the direction of gradient by subtracting the ratio 
of synapse from the weight. This ratio is corresponding to the speed and quality of 
learning. 
There are many applications in image processing and machine vision for which one can use 
this type of neural networks for processing. Traffic signs recognition could be a good 
example. Hereby, the main problem for traffic sign recognition is the design of an algorithm 
that is scale-variant and rotation-variant for recognition and classification of different 
traffic signs. This means system is not sensitive to the scale and angle of signs. During 
driving camera can see the sign from different angle of view and different scale. Therefore 
this is feature of ANN is important. 
This type of neural network has potential to learn multi-scale and multi-variant traffic 
signs [76]. This means we can train ANN for different size and different traffic sign. The use 
of artificial neural networks with their remarkable potential to derive meaning data from 
complicated/complex data and information is getting more popular in the field of image 
processing [76]. Detecting complex pattern, classification, and prediction are only few 
examples illustrating the huge potential of neural networks for image processing.

 
32 
The main advantages of using ANN in ADAS technology is flexibility, robustness, adaptivity 
in learning, real time operation after the learning phase and fault tolerance [77] [78]. It is 
known that many conventional algorithms for image processing and machine vision are 
not at all sufficiently fast in view the complexity of the tasks. It is furthermore relatively 
very difficult to define an appropriate mathematical model for solving those problems. 
ANNs however, can process the information in a similar way to the human brain. This 
means that a predefined model for solving the problems and extracting meaningful data is 
not necessary [79]. Artificial neural networks do offer many advantages. For example it 
requires less formal statistical training; another advantage is the ability of detecting highly 
complex and nonlinear relationships among independent and dependent variables. Finally, 
it has also the potential for multiple training algorithms [79]. Due to the massive 
parallelism, neural network are much faster than conventional methods. 
If the modeling of a system is not trivial and thus complex and if it is hard to formulate an 
explicit algorithm as a solution/model one should use artificial neural networks[80]. There 
is a main difference between von Neumann models for computing which are based on 
memory/processors and artificial neural networks. In an artificial neural network we are 
using a parallel architecture similar the biological brains. We do not know details of the 
complex mechanisms within the human brain; but we do know the main principles of 
neurons operations either individually or globally [81]. Another main factor of artificial 
neural networks is scalability and adaptivity in the learning phase. The system can change 
its structure based on training information. We can model very complex models and 
relation between input and output. 
Basically ANNs are appropriate for processing images [81]. However, they do have some 
weak points such as over-fitting during the training phase, their black-box nature and the 
complexity of a related hardware implementation [82].
Designing and implementing a computing system based on artificial neural networks in 
software for a high resolution image is very complex but can thereby even not be fast 
enough with regard to the real-time requirements of ADAS [80]. Therefore, we are 
interested in implementing it in hardware. There are two main problems faced by the 
hardware implementation of artificial neural networks. A first problem is the huge 

 
33 
complexity of the network due to too many connections between cells. Another challenge 
is how to design a reconfigurable architecture that can be reconfigured to execute 
different tasks [80]. To find a general purpose hardware platform and architecture concept 
capable of solving/computing different types of problems of image processing is not a 
trivial.

Download 3,22 Mb.