High-Performance OCR on Packing Boxes in Industry Based on Deep Learning

OCR is a historic but still challenging task, especially in industry conditions where it demands very high computational efficiency and accuracy. In this work, a high-performance OCR method based on deep learning is proposed. First, the region of character string is segmented using semantic segmentation network, and the tilt angle of the string is corrected so that the system is adaptive to character rotation. Then the intervals between adjacent characters are recognized by a column-classification network so that characters in the same string are well separated. Finally, each region of separated character is fed to an image-classification network to ensure a high-accuracy recognition. Different from existing networks which carry out tasks of location and classification simultaneously, in the proposed framework three different networks including the semantic segmentation network, the column classification network and the image-classification network are independent. Each of them is dedicated to its own classification task so that the classification accuracy is best. Another advantage of this framework is that it is convenient to do data augmentation. Different from traditional OCR algorithm based on deep learning, which mainly need a large number of labeled samples, the proposed algorithm only needs 100 training samples with the size of 640 * 480 to achieve an accuracy of 99.92%, moreover, the whole detection process requires only about 78 ms per image on average.