Research On Handwritten Chinese Address Recognition For Hand-held Devices

Handwritten Chinese address recognition has urgent application requirements in many fields.For example,in the logistics industry,handwritten Chinese address recognition has a bright prospect in terms of automatic collection of waybill's information and automatic sorting of parcels.Based on the long-term research and practice,researchers have made good progress on the issues of Chinese character segmentation,Chinese character recognition,Chinese text recognition,and postprocessing,etc.However,due to its unconstraint of handwriting,various and irregular address formats,handwritten Chinese address recognition is still one of the most important and challenging tasks in the field of computer vision and pattern recognition.The hand-held devices are very popular in people's daily life with an extremely rapid development in recent years.By integrating the hardware with powerful processing capability and software with wide practicability,these devices not only improve people's life,but also promote the progression of other related fields.What's more,it has created favorable conditions for the development of pattern recognition on hand-held devices.Some techniques of pattern recognition such as fingerprint recognition,ID identification,face recognition and 2D barcode recognition have been widely used in hand-held devices.Application requirements of Handwritten Chinese address recognition arise for hand-held devices.This thesis focuses on handwritten Chinese address recognition for hand-held devices,including the issues of building a compact MQDF classifier with high recognition accuracy and recognizing handwritten Chinese address with writing variations.In term of building a compact MQDF classifier,sparse coding and vector quantization technique are used to reduce the storage size of MQDF classifier to meet the requirement of hand-held device with small memory space.On the other hand,a word-level-tree framework is proposed for handwritten Chinese address recognition with writing variations,which can map all the variations of an address to its corresponding normative address format.The contributions of this thesis are as follows.1.A method to reduce the storage size of MQDF classifier using sparse coding is proposed.Two methods of sparse coding,viz.,the maximum likelihood-based method and the K-SVD method,are adopted to build two compact MQDF classifiers.The feasibility of the sparse coding method in reducing the storage size of MQDF classifier is theoretically proved,and its effectiveness is verified by experiments.2.A compact MQDF classifier developed by integrating sparse coding and vector quantization technique is presented.Based on the proposed compact MQDF classifier using sparse coding,the vector quantization technique is employed to further compress the sparse coding.This method takes full advantage of sparse coding and vector quantization technique such that it can further reduce the storage size of MQDF classifier with a little accuracy loss.The effectiveness of the proposed method is verified by experiments.3.A word-level-tree based approach is proposed for handwritten Chinese address recognition.In the segmentation-and-recognition framework,the word-level-tree is used as the storage structure of normative address format.With the recognition of address word,character index table,and some heuristic methods,all the writing variations of an address can be mapped to its corresponding normative address format.Experiments show the effectiveness of this scheme for handwritten Chinese address recognition.4.The proposed handwritten Chinese address recognition system is successfully embedded in hand-held devices.The proposed compact MQDF classifier based on sparse coding and vector quantization technique and the word-level-tree based approach are integrated to an app for hand-held devices application.This app was tested on handwritten address captured from real mail pieces and achieved promising results.