Design, integration, and evaluation of form-based handprint and OCR systems
Charles L. Wilson, Jon Geist, Michael D. Garris, Rama Chellapa · 1996
List of FiguresCompaxison of a typical error versus rejection curve with HIGH, MEDIUM, and LOW error-rate requirements 21 Example HSF form from NIST Special Database 19 27 Organization of functional components within the NIST system 28 The results of 500 HSF forms registered and ORed together 30 Results of form box removal 31 Line-bands computed from the paragraph image above 32 Segmentor results of merging components together 33 Segmentor results of splitting components apart 34 Results from processing the top paragraph image 36 Rejection versus error rates for digit, upper, and lowercase recognition between HS- FSYSl and HSFSYS2 43 Segmentation of the word "medical" by the ERIM system 45 Formation of unions from segments by the ERIM system 46 Selection of the best combination of unions of segments to represent a word from a dictionary of expected words 47 Error rate versus rejection rate for isolated digits for systems in the First OCR Systems Conference 52 Error rate versus rejection rate for isolated upper-case letters for systems in the First OCR Systems Conference 53 Error rate versus rejection rate for isolated lower-case letters for systems in the First OCR Systems Conference 54 Comparison of human error rate versus rejection rate for isolated digits with that for some system results obtained following the First OCR Systems Conference.... 55 Comparison of human error rate versus rejection rate for isolated upper case letters with that for some system results obtained following the First OCR Systems Conference 56 Comparison of human error rate versus rejection rate for isolated lower case letters with that for some system resrdts obtained following the First OCR Systems Conference 57 A reduced copy of a FAX of the form used to collect the handprint sample for the comparison of two OCR engines 60 3 21 Error rate versus rejection rate for two OCR engines in recognizing a set of 6,422 isolated images, alm ost all of which contained digits 61 22 Error rate versus rejection rate for two OCR engines in recognizing a set of 700 images, most of which contained the ten digits, one digit to a box, but about 8 type of non-character information typically found in OCR applications as described in the text 62 23 Field error rate versus rejection rate for systems in the Second OCR Systems Con- ference.See [3] for a detailed explanation 68 24 Field distance rate versus rejection rate for systems in the Second OCR Systems Conference.See [3] for a detailed explanation 69 List of Tables 1 HSFSYSl accuracies and error rates for digit fields across the first part of SD19.Part (A) reports character-level statistics and (B) reports field-level statistics 38 2 HSFSYS2 accuracies and error rates for digit fields across SD19.Part (A) reports character-level statistics and (B) reports field-level statistics.() Segmented character images from the writers in this partition were used to train the neural network classifiers.)39 3 HSFSYSl accuracy and error rates for uppercase fields across the first part of SD19.39 4 HSFSYS2 accuracy and error rates for uppercase fields across SD19.{^Segmented character images from the writers in this partition were used to train the neural network classifiers.)40 5 HSFSYSl accuracy and error rates for lowercase fields across the first part of SD19.40 6 HSFSYS2 accuracy and error rates for lowercase fields across SD19.{^Segmented character images from the writers in this partition were used to train the neural network classifiers.)41 7 HSFSYSl accuracy and error rates for Preamble fields across SD19 41 8 HSFSYS2 accuracy and error rates for Preamble fields across SD19 42 4