Algorithm of Handwritten String Segmentation Based on Recursive Rraining in Background Domain

Abstract

How to accurately cut the handwritten string especially the sticky string, has become a key part of recognizing handwritten strings. Aiming at the traditional segmentation algorithm has some problems, such as more complex､the segmentation effect is not good and so on, this paper proposes a segmentation algorithm based on background domain recursive training. This algorithm is a common algorithm for handwritten digit string segmentation and English string segmentation. The principle is the use of handwritten string at the adhesion of the background domain features and recursive neural network RNN fusion of special mechanisms. It completes by modeling, training and implementation of three steps. RNN modeling is the core of the algorithm, it contain two important parts: ①Assignment for the background domain, extracting the eigenvector value of the depression area by the principle of adjacent matching, the connection weights of the RNN input layer are calculated.② Using the minimum area selection principle to modify eigenvector values in the RNN's acceptance layer, the connection weight of the layer are calculated agin. After modeling completion, RNN is training samples､studying and remembering. Finally, use the knowledge that RNN has learned to complete real segmentation, the effect is satisfactory.