A Multi Layer Perceptron Neural Network Trained by Invasive Weed Optimization for Potato Color Image Segmentation

Abstract

Accurate recognition of external defects on potato color images is an important point in the realization of automatic computer vision-based potato grading and sorting station. Therefore, pixel-based segmentation of potato color image is an essential step in every inspection system by computer vision. Invasive Weed Optimization (IWO) is a new evolutionary algorithm which recently introduced and has a good performance in some optimization problems. IWO is a derivative-free, meta-heuristic algorithm, mimicking the ecological behavior of colonizing weeds. In this study, firstly a proper color component for potato color image segmentation using a statistical analysis on some training images is selected. Then, combining the IWO and ANN (Artificial Neural Networks) to solve pixel-based potato classification has been proposed. In this proposed algorithm, Multi Layer Perceptron (MLP) network manages the problem's constraints and IWO algorithm searches for the best network weights based on minimization of the cost function. Experimental results on more than 500 potato images show that this method can improve the performance of the traditional learning of MLP significantly.