Abstract

In this paper, we report the results obtained recently by analysing around 400 features measured from 23,000 regions segmented in 50 noisy radioscopic images of cast aluminium with defects. The extracted features are divided into two groups: geometric features (area, perimeter, height, width, roundness, Fourier descriptors, invariant moments, and other shape factors) and grey value features (mean grey value, mean gradient in the boundary, mean second derivate in the region, radiographic contrasts, invariant moments with grey value information, local variance, textural features based on the co-occurrence matrix, coefficients of the discrete Fourier transform, the Karhunen Lòeve transform and the discrete cosine transform). We propose a new contrast feature obtained from grey level profiles along straight lines crossing the segmented potential defects. Analysing the area Az under the receiver operation characteristic (ROC) curve, this feature presents the best individual detection performance yielding an area Az=0.9944. In order to make a compact pattern representation and a simple decision strategy, the number of features is reduced using feature selection approaches. The relevance of the selected features is evaluated by the calculation of a linear correlation coefficient. The results are shown in Tables of the linear correlation coefficients between features, and among features and class of defect. In addition, statistical classifiers and classifiers based on neuronal networks are implemented in order to establish decision boundaries in the space of the selected features which separate patterns (our segmented regions) belonging to two different classes (regular structure or defects).