An application of machine learning and statistics to defect detection

We present an application of Machine Learning and Statistics to the problem of distinguishing between defective and non-defective industrial workpieces, where the defect takes the form of a long and thin crack on the surface of the piece. The images of the pieces are described by means of a set of visual primitives, including the Hough transform and the Correlated Hough transform. We have compared an attribute-value learner, C4.5, a backpropagation neural network, NeuralWare Predict, and the statistical techniques linear, logistic and quadratic discriminant for the classification of pieces. Moreover, two feature sets are considered, one containing only the Hough transform and the other one containing also the Correlated Hough Transform. The results of the experiments show that C4.5 performs best for both feature sets and gives an average accuracy of 93.3 % for the first dataset and 95,9 % for the second dataset.