Two-Stage Billet Identification Number Recognition Using Label Distribution

With the progressive automation of factories, the demand for deep learning methods capable of recognizing characters is rising. A billet identification number (BIN) is a string of characters that contains all information about the billet, but it is often oriented arbitrarily. Because each plant has different features of data, it requires time and effort to secure enough data to train the model that can be applied to each plant. In addition, the existing BIN recognition model confuses characters with similar shapes when rotated because it shares a feature extractor for angle estimation and character recognition. In this study, we propose a method to solve the problems and improve the BIN recognition performance. We separate the two parts of extracting angles and characters, allowing each module to independently focus on the features of the data. Label distribution is used to enhance the angle estimation accuracy with a small dataset, and the triangular distribution results in the highest accuracy. Finally, to train rotated characters, a large amount of data that are randomly rotated is required, but by separating the angle and character module, the variation within classes is reduced, resulting in high BIN recognition performance even with a small dataset.