Training neural networks using Clonal Selection Algorithm and Particle Swarm Optimization: A comparisons for 3D object recognition

Clonal Selection Algorithm (CLONALG) and Particle Swarm Optimization (PSO) have been applied for wide spectrum of computer vision problems. However, their applications to 3D object recognition receive only little attention. In this paper, CLONALG and PSO algorithms for recognition of 3D object are discussed. Instead of using any predefined model to extract the geometrical information, the 3D object is modeled based on its 2D image appearance. Firstly, the 2D image is segmented using Otsu thresholding method. Secondly, a set of moment features that are invariant under translation, changes in scale, and rotation is extracted. Thirdly, the CLONALG and PSO are used to initialize the neural network weights. Then, the neural network training is continued by Levenberg-Marquardt algorithm. The experimental results showed that the CLONALG-LM is better than PSO-LM and the other traditional training algorithms: Levenberg-Marquardt (LM), Scaled Conjugate Gradient (SCG), Gradient Descent (GD), Gradient Descent with Momentum (GDM), Gradient Descent with Adaptive Learning Rate (GDA), and Gradient descent with momentum and adaptive learning rate (GDMA).