Neural network based defect inspection from images

In this paper, we present a vision-based method for the automatic inspection of defects on ball bearings. A supervised neural network, employing a Multilayer Perceptron trained with the BackPropagation algorithm, is applied for discriminating defects. The images of the ball bearings are first mapped to a new space, called Defect Evaluation Space, to simplify the process of analyzing the images. The obtained images are then processed in the new space and wavelet based feature vectors are extracted. Different combinations of the wavelet coefficients are used as input vectors to the neural classifier in order to carefully select those that produce the best results. The false alarm rate, evaluated on a test set of images, is low demonstrating the accuracy and robustness of the proposed approach.