Transfer learning of a deep convolutional neural network for localizing handwritten slab identification numbers

Most machine learning methods assume that previous and future data have same distribution in same feature space. This paper presents a real-world problem that violates the common assumption, and we propose a practical methodology to handle the problem. In the steel making industry, automated marking systems are widely used to inscribe slab identification numbers (SINs). In the previous work, a deep learning based algorithm was developed to automatically extract regions of printed SINs. However, as the marking system is outdated, few SINs are marked by hand in uncommon situations, and the existing algorithm does not work for the handwritten SINs. This paper proposes a practical method that uses very small training data (10 images) to localize handwritten SINs. The knowledge of mid-level layers or entire layers in the pre-trained deep convolutional neural network is transferred to overcome the shortage of training data in the target domain. Experiments were conducted with actual industrial data to demonstrate the effectiveness of the proposed algorithm.