A global learning approach for an online handwritten mathematical expression recognition system

Online flowchart understanding by combining max-margin Markov random field with grammatical analysis

“…al. [3] used a 2D dynamic programming algorithm symbol hypothesis generator, concatenating it with a Time Delayed Neural Network (TDNN) [31] symbol classifier. Besides, structure interpretation also suffers from ambiguity, which is caused by the ambiguous position of symbols and previous recognized symbols.…”

Multimodal Mathematical Expressions Recognition: Case of Speech and Handwriting

“…This is mainly done at two levels: symbol identification (segmentation and recognition) and relationships discovering (through which the identified symbols are spatially arranged). Thus, recognizing a handwritten ME includes three sequential but interdependent steps [5,6]: segmentation, symbol recognition and spatial relations interpretation. The aim of the segmentation process is to form the symbol hypotheses "h s " from the set of strokes.…”

“…Optimizing separately each step has a major drawback since the failure of one step can lead to the failure of the next one. To alleviate this problem, the simultaneous optimization of the segmentation and recognition steps is reported in various works as in [6,7]. The handwritten MER subsystem used in the architecture of Fig.…”

“…For example if we consider the case of two symbols, the relations explored including only these two symbols can be: left/right, superscript, subscript, above, under and inside. For each explored relation a cost is associated [6]. The relation which will be considered in the ME is the one having the smallest cost and satisfying the considered grammar.…”

“…It is globally based on a Time Delay Neural Network (TDNN) classifier [6]. The corresponding output is a list of Nbest classes with their scores which are normalized in the range [0, 1].…”