Efficient search strategy in structural analysis for handwritten mathematical expression recognition

Problems in local ambiguities in handwritten mathematical expressions are often resolved at the global level. For a well performing recognizer, multiple local hypotheses should be kept as long as possible until the ambiguities are resolved by a global analysis. We propose a layered search framework for handwritten mathematical expression (ME) recognition. From given handwritten input strokes, ME structures are constructed through adding a symbol hypothesis one by one, considering every possible symbol identity and spatial relationship with the existing structure. A cost reflecting the likelihood of a structure is estimated for each newly expanded layer so that a bestfirst search algorithm is applied to seek the most likely structure. The elegance of our method is in that while all the possibilities are examined, the search complexity is made manageable by applying admissible heuristics. Further complexity reduction is achieved by delaying the symbol identity decision. Unless a symbol identity causes structural alternatives for the remaining input strokes, the identity can be determined after the complete structure is fixed. Such a delayed decision reduces undesirable search space expansion. In an implementation targeting high school level MEs, our method achieved high speed with a high level of accuracy which resulted from the system’s capacity to examine a large number of possibilities.