Evaluation of Surface Roughness by Machine Vision Using Neural Networks Approach

Abstract

Surface quality of industrial components is critical from operational, ergonomics and esthetics point-of-view. Surface roughness measurement using traditional contact-type instruments may not be feasible in the industries insisting on 100% inspection and monitoring. Machine vision-based machine learning has a potential of facilitating automated inspection of manufactured components for their surface quality. The paper presents a machine vision-based machine learning approach that works on the principle of surface texture characterization by vision-based texture analysis techniques followed by supervised machine learning using multilayer feedforward artificial neural network with backpropagation for fitting the response (surface roughness) with the inputs (vision-based texture parameters). Performance of various texture analysis techniques based on the histogram, gray level co-occurrence matrix, Fourier and wavelet transform used for generating the training data and training algorithms used for training the networks are compared. The approach can be potentially used to estimate surface roughness of industrial components.