There are many mathematical expressions (MEs) in the science and technology documents, yet the OCR in commerce couldn’t efficiently understand the ME in these documents. We, in this paper, focus our research on some key technique of automatic ME recognition, such as ME identification in Chinese document image, binary algorithm for ME image, symbols segmentation/recognition in ME and structure analysis for ME. The main contributions of this thesis include: (1) A ME identification algorithm is proposed. Based on Chinese character recognition and ME symbol recognition, Chinese character and non-Chinese character are distinguished (if there are Chinese characters in the document image). Then according to some features of ME symbols, ME symbols are extracted from non-Character symbols. Finally, using format information, isolated ME is discriminated. ME identification accuracy with 91.19% on the database with 148 document images which contains 3690 MEs is reached. (2) A binary algorithm for ME image is proposed. On one hand, a binary algorithm based on connect component is adopted in order to decrease broken probability of ME symbols. On the other hand, a binary algorithm based on histogram is adopted so that touching probabitlity among adjacent symbols is reduced. The two binary algorithms are integrated based on symbol recognition. (3) A symbol segmentation algorithm in the ME image based on three-stage dynamic programming (DP) is introduced. DP algorithm is firstly adopted to segment sub-images in vertical direction. Then symbols in every block are segmented using DP algorithm in horizontal direction. Finally, broken symbols are combined based on DP algorithm. The experiments were implemented on a database with 1322 images and the symbols segmentation accuracy reached 96.40%. (4) A symbol recognition method with non-symbols rejection model is proposed. Symbol recognition accuracy with 98.58% on the database was obtained. (5) A hierarchical structure analysis algorithm for ME is proposed through reconstructing the ME global structure. The method decomposes the ME into several basic sub-expressions, which efficiently decreases ME structure analysis complexity. ME structure analysis accuracy on the database with 1322 ME images reaches 87.59%. (6) An automatic understanding system for ME is built. The expereiment datebase is consisted of 148 document images which contain 3690 MEs. The accuracy of 81.24% in ME recognition was obtained. Now, the ME recognition system has been integrated into HWOCR, and it has been in commerce.
修改评论