Intellectual two-sided card copy

The paper is devoted to intellectual algorithm for processing of scanned images of various two-sided cards such as ID, bank, business cards, etc. Card images significantly differ from common document images and conventional approaches developed for text document image processing do not operate well on card images. Our technique provides a sophisticated arrangement of both images of a card, front and back, for single-page and duplex printing. The method detects location, skew angle and orientation of card image. Orientation detection is based on Arabic numerals segmentation and recognition. Recognition of numerals on card images is a challenging problem, since often cards have complex color background; digits can have arbitrary color and typeface. For digits segmentation on complex background we apply labeling of connected edge regions, whereas edge pixels are labeled for dark numerals on light background and vice versa independently from each other. We construct a decision tree with a great generalization capability for recognition of each numeral; it improves recognition for various typefaces. We propose a final decision rule for image orientation detection based on recognition outcomes. Our algorithm operates very fast and outperforms OCR applications ReadIRIS and FineReader in numerals recognition and orientation detection for card images. Also our algorithm works well enough for majority of originals.