You are welcome to participate in the Chinese Handwriting Recognition Competition 2011, which is held with the 11th International Conference on Document Analysis and Recognition (ICDAR2011), Beijing Friendship Hotel, 18-21 September, 2011. The competition is aimed to test the state-of-the-art technology in online and offline handwritten Chinese character recognition and handwritten Chinese text recognition, to find the merits and insufficiencies of current methods, and to motivate further research. All researchers and students from the academia and the industry can participate. We will publicize free training data and trial test data. The participants are required to submit their executable systems, which will be evaluated on preserved test data. The evaluation results will be summarized in a paper to be included in the proceedings of ICDAR2011, and will be presented at the conference. The winners of competition tasks will be presented certificates at the conference.
The Chinese Handwriting Recognition Competition has four tasks:
1. Offline Chinese Character Recognition (Task 1)
2. Online Chinese Character Recognition (Task 2)
3. Offline Handwritten Text Recognition (Task 3)
4. Online Handwritten Text Recognition (Task 4)
Participations in any one or multiple tasks are welcome. Performance evaluation and system ranking will be based on single task, i.e., each task will be evaluated separately.
For isolated character recognition (Task 1 and Task 2), the character set is confined as the set of 3,755 Chinese characters (level-1 set of GB2312-80), which is often tested in Chinese character recognition research. The participating systems will be ranked in the character recognition accuracies. Additional statistics of program size (including the storage size of classifier parameters) and recognition speed on a standard computing platform will be reported as well.
For handwritten text recognition (Task 3 and Task 4), we will provide text page images (in the case of online recognition, page ink data) with the text lines segmented, such that the participants do not have to deal with text line segmentation. The participating systems will be evaluated in respect of average character-level correct rate. The number of correctly recognized characters will be counted by minimizing the string edit distance between the output text and the ground-truth text of each text line. The systems will be ranked in the character-level correct rate, and additional statistics of program size and recognition speed will be reported.