A multi-level perception approach to reading cursive script

摘要: Abstract Reading cursive script is the problem of transforming language from iconic form human handwriting to one ordinal representation. The involves elements visual perception at level processing and those understanding a higher level. approached as multi-level in which word image transformed through representational hierarchy. levels are based on descriptions that use points, contours, features, letters, words. Global control hierarchical until an intermediate after it heterarchical. A feature representation generated bottom-up using statistical dependencies between letters features. Ratings for partially formed words computed stack lexicon represented trie. Several heuristics low- intermediate-level introduced, including: reference-line finding projection profile analysis, letter segmentation local lower contour minima areas with low vertical profiles, simultaneous encoding contours their topological relationships, extracting features (e.g., middle loop, upper-zone stroke), shape-oriented events. Two modes learning defined: initial training user feedback unsupervised adaptation writer. Experiments demonstrating promise approach described.