Localization of slab identification numbers using deep learning
Sang Jun Lee, Jaepil Ban, Hyeyeon Choi, Sang Woo Kim · 2016
In the steel industries, recognizing product information is an important task for the management of the manufacturing processes. For real factory scenes, localization of product identification numbers is conducted prior to recognition to obtain a satisfactory performance. The objective of this paper is localization of slab identification numbers in real factory scenes. Traditionally, most researches in the field of image processing and pattern recognition were focused on feature representation or shallow learning. However, conventional rule-based algorithms heavily depend on carefully engineered feature values and require heuristic parameter tuning. To overcome these limitations, a deep learning based algorithm is proposed for the localization with the minimum of manual interventions. This paper contains construction of training data, labeling process, and an architecture of a deep convolutional neural network. The performance error is remarkably reduced to 2.19% by the proposed algorithm compared to 4.59% in the previous work. By using a data-based method, this algorithm is easily expandable to apply for other applications.