The project was made within ASIC Art for an Israeli client: Elbit Vision Systems. The percent of correct recognition on the test set of images (other than the one "learned") was 89% (for 5 kinds of different defects). The solution was a result of my research because classical solutions didn't give good results (about 50% of good recognition).

The solution comprises a preprocessing stage and a neural network based-process. The neural network used was a Multilayer Perceptron trained with Backpropagation algorithm. In the preprocessing stage the main features that characterize the defects are extracted from the original images. It is almost a rule the using of Discrete Cosine Transform in order to extract the features of an image, this preprocessing task being a under-optimal transform after Karhunen-Loeve. This transform didn't give good results, so I had used combined methods histogram manipulation, Gabor filtering, statistic moments. The later method allowed us to increase the percent of good recognition. The pre-processing is an issue under development/research.