Adaptive confidence transform based classifier combination for Chinese character recognition
Introduction
Off-line Chinese character recognition remains a difficult pattern recognition problem although much research has been done (Hildebrandt and Liu, 1993). One effective approach to improving the performance is to combine multiple classifiers (Xu et al., 1992). This strategy is quite popular in handwritten alphanumeric recognition. Some researchers have paid attention to the multi-expert approach in Chinese character recognition (Guo et al., 1996), but the proposed methods are usually limited to their own specific features and classifiers.
In Chinese character recognition, the number of categories is much larger than in alpha numeric recognition. At least 3755 classes of Chinese characters are to be handled. This is in sharp contrast to alpha numeric recognition, in which there are less than 100 categories. This difference gives rise to a series of differences between the two kinds of recognition and makes it more difficult to combine multiple classifiers in Chinese character recognition.
Xu et al. (1992) grouped combination methods into three types according to the information used: abstract level, rank level and measurement level. For alphnumeric recognition, it is easy to devise several classifiers (e.g., Suen et al. (1992) used up to four classifiers for handwritten numeral recognition). As a result, abstract level information-based combination such as majority vote (Battiti and Colla, 1994) and rank level information-based methods such as the Borda count method (Ho et al., 1994) can achieve satisfactory results in such recognition tasks. However, in Chinese character recognition usually there are only two or three classifiers available because of the much greater complexity in the design and training of classifiers and more importantly, the lack of high-performance recognition algorithms. Consequently, the abstract level or rank level information seems insufficient for effective combination. So it is imperative to take full advantage of measurement information.
In alpha numeric recognition, neural network based classifiers (especially BP nets) are widely and successfully employed. Because the outputs are good estimates of a posteriori probabilities (Richard and Lippmann, 1991), these outputs can be directly used in combination (Battiti and Colla, 1994). For Chinese character recognition, things are quite different. The large number of classes makes it extremely difficult to train proper neural networks. So most Chinese character recognition systems adopt traditional feature vector matching for discrimination. The immediate outputs of the classifiers are usually in the form of distances rather than probabilities.
Fully parallel combination, in which an input pattern passes all recognition engines, is widely used in alpha numeric recognition because the computing cost of individual classifiers is relatively low. In Chinese character recognition the computing load of a single algorithm is already very heavy and a fully parallel approach will make the overall speed too slow to be acceptable. So how to accelerate the processing is also a problem we must face.
The above three problems, especially the second one, are the main concerns of this paper and distinguish our work from previous research in the field of multiple classifier combination.
This paper is organized as follows. In Section 2an adaptive confidence transform (ACT) is proposed to estimate a posteriori possibilities from raw measurement values. In Section 3consensus theory is used to make the final decision. Section 4is devoted to a reliability-based scheme that can increase the processing speed. Section 5presents the specific goals we want to achieve through multi-expert combination in Chinese character recognition. Experimental results are shown in Section 6. Section 7is a summary.
Section snippets
Adaptive confidence transform (ACT)
In theory, measurement values can provide much information for combination. However, in order to utilize such information effectively two problems must be handled. First, the raw measurement values such as distances between feature vectors are not a good indicator of recognition reliability. Second, the confidence measurements of different recognition methods are incomparable (Ho et al., 1994). To solve the two problems, an adaptive confidence transform (ACT) to convert distance measurement to
Consensus theoretic combination
Through ACT, the outputs of different classifiers are all converted to a posteriori probabilities. Then another question arises: how to combine these probabilities. Huang and Suen (1994) proposed the Linear Confidence Accumulation (LCA) method which sums up all the confidence values. This implies equal weights for every classifiers and when the performance is very different from one classifier to another the result will not be satisfactory. For alpha numeric recognition, we can just input these
Recognition reliability-based speedup scheme
The combination process described in 2 Adaptive confidence transform (ACT), 3 Combination of the a posteriori probabilities of different classifiersis illustrated in Fig. 3. It is fully parallel in the sense that every sample must pass all the classifiers. As mentioned before, speed requirement poses an extra challenge in Chinese character recognition. In this section a recognition confidence based speedup scheme will be presented to accelerate the processing.
Suppose that there are M
Application in Chinese character recognition
Off-line Chinese character recognition can be divided to two groups: printed Chinese character recognition (PCCR) and handwritten Chinese character recognition (HCCR). For both recognition tasks the general motivation of classifier combination is to improve performance. However, PCCR and HCCR are at different stages of development. So the specific objectives we hope to achieve through combination are not the same for the two problems.
PCCR is a relatively simple pattern recognition problem. It
Experimental results
In this section the experimental results are presented to show that through combination the PCCR becomes more robust and HCRR's recognition rate is increased despite the wide gap between the performance of the two methods.
Conclusions
Nowadays, the combination of complementary classifiers is a subject of intense research interest in pattern recognition, especially in the field of character recognition. As mentioned before, combination possesses certain characteristics when applied to Chinese character recognition. Focusing on these characteristics this paper proposes an ACT-based combination method, applies it to PCCR as well as handwritten character recognition and achieves promising results.
We conclude by indicating the
Acknowledgements
We would like to thank Dr. Youbin Chen and Dr. Hong Guo, who provide us the individual recognition algorithms. In addition, Ms. Liangrui Peng helps us with tests on printed Chinese character recognition. Chinese Hi-tech 863 Plan (Project No: 863-306-03-05-6) and National Science Foundation (Project No: 69682003) have also supported our research. Finally, we would like to thank the reviewers of this paper for their valuable comments and suggestions.
References (18)
Recognition technology frontiers
Pattern Recognition Letters
(1993)- et al.
Democracy in neural nets: Voting schemes for classification
Neural Networks
(1994) - et al.
Optical recognition of handwritten Chinese characters: Advances since 1980
Pattern Recognition
(1993) - et al.
Consensus theoretic classification methods
IEEE Trans. Systems Man Cybernet.
(1992) - Chen, Y., 1997a. Off-line handwritten chinese character recognition based on crossing line feature. Proceedings of...
- Chen, Y., 1997b. Research on handprinted Chinese character recognition. Ph.D. Dissertation, Tsinghua University,...
- Guo, H., 1997. Improve the robustness of printed Chinese character recognition. Ph.D. Dissertation, Tsinghua...
- Guo, H. et al., 1996. Comprehensive recognition method for improving the robustness of recognition of printed chinese...
- et al.
Decision combination in multiple classifier systems
IEEE Trans. Pattern Anal. Machine Intelligence
(1994)