Open-World classification problem in loyalty cards image classification task.

Classical machine learning algorithms make a closed world assumption that the classes represented in the test sample should also be represented in the training sample. However, when working with the real world, the model should be able to distinguish the types of objects that she saw before (in which she studied) from the types that she had not previously met. This learning paradigm is called the open world classification. Recently, machine learning products such as smart assistants or automatic cars that face the problem of identifying unknown classes in the real world have become increasingly popular, so the need to classify the open world is becoming increasingly important. In this paper, we will solve the open world problem in the task of classifying images of loyalty cards using the example of Cardsmobile. We will train models with different architectures and compare their results. We will also consider how to create an ensemble of models for the task of classifying an open world and show that when using an ensemble, the quality increases significantly.