Takuma Akagi

A new approach for multilingual address recognition

In this paper, we proposed, a new method, for multilingual address recognition which is based, on MAP (maximum a posteriori probability) estimation. In the method, an input document image is decomposed to fundamental factors, which are characters, words, text lines and, address areas. These factors are constructed hierarchically in accordance with the degree of their abstraction. As a statistical method is applied, to the formulated relationships among the factors, recognition of a document image is essentially equal to finding the proper position of factors, and classifying each character is attained, through a learning process involving samples of any language type, and thus proper weight coefficients are found, without heuristic management. We constructed, a trial recognition system based on the proposed method. Experimental results show that the proposed system is very promising not only for Japanese envelopes but also for English ones. The proposed method, of address reading is applicable to any type of language, i.e., it is multilingual, by virtue of proper learning from samples.

An A Posteriori Probability Calculation Method for Analytic Word Recognition Applicable to Address Recognition

In this paper, we propose a novel calculation method of an “a posteriori” probability for analytic word recognition.The method is suitable for address recognition tasks. Our previous method needs calculation over all words in a lexicon,while the proposed method only needs calculation on the concerned word. In the address recognition task, lexicon size becomes very large and only a part of it can be handled. Hence, the previous method cannot be used, and the proposed method is convenient. Experimental results show that the proposed method guarantees high precision of probability calculation.

Bayesian Best-First Search for Pattern Recognition - Application to Address Recognition

In this paper, we propose a novel algorithm “Bayesian Best-First Search (BB Search)”, for use in search problems in pattern recognition, such as address recognition.BB search uses “a posteriori” probability for the evaluation value in best-first search. BB search is more flexible to changing time limits compared to beam search used in conventional pattern recognition approach. It does not need designing a heuristic function for each problem like A* search.We demonstrated a 12.4% improvement over beam search on an address recognition experiment using real postal images.