Video Segmentation Using Neural Network and Distributed Genetic Algorithm

The paper presents a neural network and distributive genetic algorithm (DGA) based segmentation method which can extract and track moving objects in video. Each pixel in a frame is labeled and then this labeled configuration is modeled by MRF (Markovian random field) technique in accordance with the correlation between pixels of a frame across spatio-temporal domain. DGA is used for optimization due to its effective exploitation within search space. Chromosomes are not evolved randomly but based on the results of previous frames and only unstable chromosomes are evolved using crossover and mutation. These mutually exclusive labeled regions are fed to an unsupervised neural network which classify these regions into background and foreground with the help of sigmoid function. Each mutual exclusive region is separately modeled by neural network and wherein each pixel is modeled by two hidden layer neurons results in classification of that region of which pixels belong into either foreground or background. The parameters for hidden layer are obtained from subtraction of frames. Results are shown which confirms the efficiency of the method on both ends, resources available and time.