Incremental Training of Convolutional Neural Networks for Recognition of Marking Symbols

The paper solves the problem of classifying marking symbols on a metal surface using neural networks that incrementally relearn in the process of operation. A set of images of railway wheel marking symbols is used as source data. An incremental method of training various convolutional neural networks is investigated to identify the most successful models with the least forgetting of old knowledge. A description of experiments to assess the accuracy of neural networks is given. A model was selected that demonstrates the highest accuracy in classifying marking symbols when additional training of neural networks using the incremental method.