Evolutionary Algorithm for Landmark Detection

From Wikipedia, the free encyclopedia
Jump to: navigation, search

there are several algorithms for locating landmarks in images such as satellite maps, medical images etc.
nowadays evolutionary algorithms such as particle swarm optimization are so useful to perform this task. evolutionary algorithms generally have two phase, training and test.

in the training phase, we try to learn the algorithm to locate landmark correctly. this phase performs in some iterations and finally in the last iteration we hope to obtain a system that can locate the landmark, correctly. in the particle swarm optimization there are some particles that search for the landmark. each particle uses a specific formula in each iteration to optimizes the landmark detecting.

The fundamental particle swarm optimization algorithm used in training phase generally as follows:

Randomly initialise 100 individuals in the search space in the range [-1,1]
LOOP UNTIL 100 iterations performed OR detection error of gbest is 0%
FOR each particle p
Detection errors at x = 0
FOR each image i in training set
FOR each pixel coordinate c in i
Evaluate x of p on visual features at c
IF evaluation is highest so far for i THEN
Detected position in i = c
IF distance between detected position and marked-up position > 2mm THEN
Detection errors at x = Detection errors at x + 1
Fitness of p at x = 1- ( Detection errors at x /Total no. of images in training set)
IF new _tness of p at x > previous _tness of p at pbest THEN
pbest _tness of p = new _tness of p at x
pbest position of p = x of p

IF new _tness of p at x > previous gbest _tness THEN
gbest _tness = new _tness of p at x
gbest position of p = x of p
FOR each particle p
Calculate v of p
IF magnitude of v > v max THEN
Magnitude of v = v max
Move x of p to next position using v
IF x of p outside [-1,1] range THEN
x of p = -1 or 1 as appropriate
Output gbest of last iteration as trained detector d