Training method of deep learning models for ordinal classification using triplet-based loss and training apparatus thereof

摘要

The present invention relates to an image processing technology using machine learning. According to the present invention, the deep learning model training method for an ordinal classification problem comprises: a step of forming a convolutional neural network (CNN) which uses a learning target as input and includes a branch point and two end points divided at the branch point to generate classification loss and triplet loss; a step of calculating the classification loss for end-to-end training; a step of calculating the triplet loss enabling the network to learn ordinal characteristics; and a step of performing relative triplet sampling based on the calculated classification and triplet losses to update the final loss value to the network. Accordingly, the method can realize effective training and loss control.