Flaw Classification by Using Artificial Neural Network and Wavelet

Abstract

This paper presents a methodology to 2-D flaw-shape recognition by combining a neural network and the wavelet feature extractor. This approach consists of three stages. First, the 2-D pattern of an object is retrieved from image and then transformed to complex contour, which is described by the coordinates of its shape. Then, feature extraction is performed to this contour representation. Fourier descriptor (FD), principal component analysis (PCA) and wavelet descriptor (WD) are employed in this stage, and their performances are compared and discussed. In the third stage, artificial neural networks, including two different types of multi-layer perceptron (MLP) and Kohonen self-organizing network, are used as the classifier based on the feature sets extracted in the second stage. The numerical experiments performed on the recognition of simulated shapes demonstrate the superiority of the WD feature extractor (both used for MLP and Kohonen network classifiers) to the other two: PCA and FD, especially when the raw data have poor signal-to-noise ratio (SNR). The application to the real ultrasonic C-scan image flaw-shape classification shows the effectiveness of the proposed approach to the field of PVP.