Evaluation of multilayer perceptron and self-organizing map neural network topologies applied on microstructure segmentation from metallographic images

Abstract

Artificial neuronal networks have been used intensively in many domains to accomplish different computational tasks. One of these tasks is the segmentation of objects in images, like to segment microstructures from metallographic images, and for that goal several network topologies were proposed. This paper presents a comparative analysis between multilayer perceptron and self-organizing map topologies applied to segment microstructures from metallographic images. The multilayer perceptron neural network training was based on the backpropagation algorithm, that is a supervised training algorithm, and the self-organizing map neural network was based on the Kohonen algorithm, being thus an unsupervised network. Sixty samples of cast irons were considered for experimental comparison and the results obtained by multilayer perceptron neural network were very similar to the ones resultant by visual human inspection. However, the results obtained by self-organizing map neural network were not so good. Indeed, multilayer perceptron neural network always segmented efficiently the microstructures of samples in analysis, what did not occur when self-organizing map neural network was considered. From the experiments done, we can conclude that multilayer perceptron network is an adequate tool to be used in Material Science fields to accomplish microstructural analysis from metallographic images in a fully automatic and accurate manner.