Learning defect classifiers for visual inspection images by neuro-evolution using weakly labelled training data