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

Download 3,22 Mb. Pdf ko'rish
bet	32/81
Sana	16.05.2024
Hajmi	3,22 Mb.
	#238917

1 ... 28 29 30 31 32 33 34 35 ... 81

Bog'liq
Alireza Fasih

5.2.8 Fitness function in the genetic algorithm

5.2.7

Crossover and mutation in the genetic algorithms
After selecting parents from the population by any suitable mechanism, the genetic
algorithm operator crossover is applied on the selected parents. Crossover breeds the
selected parents to produce new children for the next generation. For the reproduction
phase and to produce the next generation, breeding the parents to produce new children is
necessary, otherwise the evolution process cannot proceed to better solutions. Crossover
can (but not every time) produce children that have better fitness than the parents. We use
a 2-
point crossover in which two crossing sites are selected in the parent’s chromosomes
randomly and the genes between the crossing sites are interchanged between the two
parents. Mutation is a genetic operator that maintains genetic diversity from one
generation of a population to the next. The purpose of mutation is to prevent trapping into
local minima. In our approach of genetic algorithm that uses real number chromosomes,
any arbitrary number (gene) in a randomly selected parent is changed by a randomly
selected number that falls in the interval to which all the chromosomes genes belong.
5.2.8

Fitness function in the genetic algorithm
Fitness function plays an important role in determining the exact solution. It determines
the fitness for every chromosome in every generation by comparing it with the original
solution. The exact solution can be reached when an exact (or near to exact) match is found
between a chromosome and the original solution. The search may also finish when a
specified number of populations/iterations has been completed. The fitness function used
to compare every output image with the target image in our genetic algorithm approach is
based on Euclidean distance between the two images. It first calculates a cost which is a
measure of to which extent two images differ from each other.
This cost value is then mapped to a fitness value which represents the fitness of the output
image. Throughout the evolution process, the genetic process aims to minimize the cost
function and increase the fitness value.

48
(5-3)
𝐶𝑜𝑠𝑡 (𝑖, 𝑗) =
∑
∑
(𝐼(𝑖, 𝑗) − 𝑇(𝑖, 𝑗))
4𝑀𝑁
(5-4)
𝐹𝑖𝑡𝑛𝑒𝑠𝑠𝑉𝑎𝑙𝑢𝑒 = 1 − 𝐶𝑜𝑠𝑡
In Equation 5-3, I represents an M×N input image into the CNN processor and
T
does
represent a M × N target or reference image. The Euclidean distance is useful to find the
pixelwise difference between the input and the target image. The denominator 4×M×N is
used for normalizing the cost between 0 and 1. The input image is normalized in the range
[-1, +1] before being fed into the CNN.

Download 3,22 Mb.

1 ... 28 29 30 31 32 33 34 35 ... 81

Download 3,22 Mb.

Pdf ko'rish