Application Of Max-margin Markov Random Field In The Recognition Of Handwritten Flowcharts

Flowchart is a widely applied format of recording, and the research of flowchart’s recognition is becoming focused. As a two-dimensional handwriting language, the main difficulty of recognition lies in symbol ambiguity and syntactic ambiguity. The origination of this ambiguity is the contradiction of error propagation in symbol segmentation, symbol recognition and syntactic analysis. In this paper, we are seeking a global method which will both tackle symbol ambiguity and self-adapt to ascents of flowcharts. Max-margin Markov Random Field is such an undirected graphical model that can capture relationship between random variables. We wish M3 N would automatically learn syntactic information from data to improve recognition performance. The main contribution of this thesis are the following:1. Learning M3 N by structured SVM. Structured SVM is a heuristic searching algorithm that can find max-margin cutting plane in linear-time complexity. This solves the learning difficulty encountered in CRF for loopy graphical model. Besides, this is the first time structured SVM is applied in two dimensional handwritten recognition.2. The building of high order, multi-variable M3 N solved symbol segmentation and recognition simultaneously. Our model labels stroke and stroke relationship at the same time and groups strokes according to stroke relationships. No symbol hypothesis is needed. 3. Features are obtained by refining raw features. Feature refiners are classifiers that classify items according to flowchart’s ascent. This achieved the property that the recognition is independent of flowchart’s ascent.4. The building of M3 N on symbol level labeled symbols’ relationships, which give structure description of flowchart.Our experiments on two different flowchart dataset proved that our high order, multi-variable M3 N has achieved state-of-the-art performance. Our model proved its elegancy and usefulness.