Using OCR for Census Data Capture in China National Bureau of Statistics of China

Background 5 population censuses have been conducted in 1953, 1964, 1982, 1990, 2000 respectively 1953,1964 census: manual tabulation Since 1982 census, using computer for data process. 1982,1990 census, manual data entry 2000 census, using OCR for data capture 2006, the second Agriculture census: also use OCR for data capture 2000 population census and 2006 agriculture census are the two cases of OCR use for large- volume data capture

Two cases of OCR for large-volume Census Data Capture The data capture of 2000 Population Census  Census reference time: Nov. 1,2000  Data Capture cycle ： Jan. -- June., 2001.(6 months)  Scale ：  Types of Census Form ： 4 Short Form: 49 Census items, 90% HH, 360 million A4 size double sheets in total Long Form: 95 Census items, 10% HH, 40 million A3 double sheets in total Other Forms: Death pop, temporary residents. 10 million A4 double sheets in total  Original Census data ： About 64 GB  Image volume: 5.5TB

The Second National Agricultural Census  Census reference time: end of 2006  Data Capture cycle ： April to mid-July, 2007,100 days  Scale ：  Types of Census Form ： 8  Total census items ： 541  Total agricultural Families ： 250 millions  Total Census Forms ： about 500 million pieces of paper  Original Census data ： about 300GB  Image data ： 40TB Two cases of OCR for large-volumes Census Data Capture

Organizational Structure for data process EA Town County 113 Checked & packed Checked & packed Coding Prefecture 5 million Data Capture editing Province 31 NBS Data process Data capture was decentralized at prefecture offices Village 0.9 million Checked & packed

Function framework of OCR data capture (2006 agriculture census) task management System management scanning Editing and checkup Data management Image management User managementSystem Initialization Log management Space management Sys management Address base Client management Archiving management process management Check scan numeric data Chinese character numeric data edit OCR Image reported Restore Delete EnquiryBrowse Importexport Forms check Scanner self-inspection Browse Input Restore Output Backup Delete QA Add scan Alternative scan Batch form scan Generation image management ID Image merger Chinese character edit Generation census form management ID Repeat scan English character Special character Progress monitoring Receiving file Statistics summary Information display System Functions English character edit Special character edit checkup

The scanning module generates image files and transmits them to image management module and also transmits the status information to task management module. The task management module executes task distribution according to the state of vacancy of each OCR clients. The Process of OCR data capture

The OCR module performs recognition of numerical data and Chinese characters and transmits the data and Chinese characters to data management module and transmits the status information to task management module. The Process of OCR data capture

The task management dispatches the data to edit module for editing. If original image is needed, corresponding image is fetched by image management module for comparison, the cleansed data after edit are returned back to data management module. when data capture work is all finished, report upward the data. The Process of OCR data capture

Quality Control To ensure the quality of captured data, quality control is executed in three stages: scanning, recognizing and data editing. During the process of scanning, recognizing batch cover data and scanner count, the system checks if the total page count, total household count for each batch are consistent with the results of scanning; Comparing the actual address code with address code repository, ensure that the address codes are validity, uniqueness and correctness. During the recognition, collecting real time statistics for rejection ratio and suspect ratio. If rejection ratio and suspect ratio is too high, the task administrator checks the reason.

During the process of editing, checking the consistency between recognized record count and the record count in controller document; Checking the basic logic relationship and value range; indicate the items which have mistakes in logic relationships or value ranges, recognition results and corresponding items from original scanned images are displayed comparatively in parallel windows, and convenient modification means are provided for those which need get modified. After the whole set of data has been captured, quality is assured through executing sampling quality check through all phases Quality Control

Main Problems and Solutions In large-scale census data capture projects, there’re three aspects of problems we regard as the most outstanding: 1. How to enhance OCR’s recognition capability. 2. Availability and reliability of the system. 3. Project management. What we have done are: 1. Improve the capability of recognizing numeric characters Two kinds of recognition algorithms and two kinds of recognition engines based on the two algorithms were developed, after a series of onsite test, which better suites the census project is chosen. 2. Improve the recognition capability for Chinese characters By collecting large number of actual samples and training the recognizer, recognition capability for Chinese names is improved.

Main Problems and Solutions 3. Improve orientation capability Aiming at print deviation and filling deviation, smart locating algorithm has been developed which has minimized the impact of the print deviation and filling deviation. 4. Enhance efficiency of recognition Improve the fundamental software of scanner, to achieve the best match between hardware drivers and OCR software and improve the efficiency of recognition. 5. Improve the quality of forms filling Prescribe the filling standards for form filling so that OCR error rate will be reduced, meanwhile rejection rate could also be reduced.

Main Problems and Solutions 6. Establish regulation, working guidance and processes to make every data entry site to execute work following uniform regulations, processes and standards. 7. Strengthen the training. we organized centralized training and on-site training for the users. Lecturing and actual operations are combined during centralized training, through the combination of these two ways, the familiarity with the system has get deepened. 8. Organize multi-target pilot. We organized multiple pilots in many locations aiming at different targets.

Lessons Learned Using advanced technology to raise efficiency Combining technical and administrative methods to resolve quality problems and security issues Choose partners with the higher capability of system development and service Early project preparation Manage project with partners Training, pilot projects and management is the key to success Control the printing quality of the census forms and census data filling quality Project change control

Prospect of the 2010 Population Census Census time: Nov. 1, 2010  Short form and long form, death population form  Foreigners living in China are considered to be enumerated Data capture in 2011  OCR data capture will be the main data entry method  Modifying the existing system of agricultural census and make some innovate  Adding more OCR equipments

THANKS