Neural network training for automated defect detection in additive production

The article discusses the methods of training your own convolutional neural network, with the aim of detecting non-standard objects, in particular, typical defects that arise in additive manufacturing when printing by layer-by-layer fused technology. The features of the choice of the optimal framework are considered, the elements of the mathematical model of the convolutional neural network are described, the features of installing and configuring the library are shown, the technique for retraining and training a new model is described, the features of annotating the initial images for training the model, the numerical indicators obtained for the speed and accuracy of the model are shown. The prospects for the development of the system and the ways of improving the proposed method are shown.