Loading...
Please wait, while we are loading the content...
Similar Documents
Experiments in the recognition of hand-printed text : Part I-Character r
| Content Provider | Semantic Scholar |
|---|---|
| Author | Ecognition |
| Copyright Year | 2010 |
| Abstract | Among the many subject areas in the field of pattern recognition, the recognition of machine-printed and hand-printed alphan:umeric characters has perhaps been the classic example to which people have referred in exemplifying the field. Interest in character recognitio.n has long run high; an extensive literature in handprinted character recognition alone dates back to at least 1955.136 In recent years, the .recognition of machine printing has become a commercial reality. Following the introduction of the highly controlled E13B magnetic font by the banking industry, several advances in optical character recognition (OCR) capability have been brought to the marketplace. The trend of these advances is toward the acceptance of broader and less controlled classes of input: from single, stylized fonts to multi-font capability; from high-quality copy to ordinary inked-ribbon impressions, and even to multipart carbons of surprisingly poor quality. Still, in contrast to hand printing, the approaches to OCR have been able to rely on the lack of gross spatial distortions in the character images, and to make considerable use of templates. Progress in the off-line recognition of hand printing has been slower. The problem is intrinsically harder than that of OCR, as reflected in the fact that the human recognition error rate for isloated, hand-printed characters is many times higher than for machine printing. The great spatial variability of hand-printed characters has led many researchers to explore nontemplate methods for recognition. Thus, the maj or effort of many researchers has been the exploration of unique methods of preprocessing, or feature extraction, applied to the hand-printed char~cter ~mages. Dinneen,1 in one of the earliest papers, InvestIgated local averaging and smoothing operations to improve the quality of the character image. Similar . operations have appeared as a part of many other approaches. 4,7 Lewis,16 Uyehara,21 Stern and Shen,23 and Rabinow Electronics31 have used schemes in which the sequence of intersections of a slit scan with the character image, or the equivalent, gave rise to features for classification. Lewis15 . was one of the relatively few to emphasize the use of multiple-valued rather than binary-valued features, an ingredient we have found important in our own work. Singer12 and Minneman30 employed a circular raster, which can facilitate size normalization and rotation invariance. Unger,7 Doyle,9 and Glucksman27 have emphasized features derived from shape attributes such as lakes, bays, and profiles. The building up of a character representation from component elements matched to the image, such as short line segments or portions of the boundary, has been attempted by Bomba,4 Grimsdale et al.,6 Kuhl,19 and Spinrad.26 Correlation techniques have been tried by Highleyman13 and Minneman. 30 Contour-following with a captive flying-spot scan or its simulated equivalent has appeared in the work of Greanias et al.,20 Bradshaw,22 and Clemens.28 The work of Greanias et al.,20 is especially significant because it led to the method used in the IBM 1287 character reader. Other 'workers have placed greater relative emphasis on classification techniques and on the selection of features from a feature set or pool. Chow16 ,29 has long worked with statistical classification methods. Bledsoe and Browning3 and Roberts8 applied adaptive procedures to features obtained from more or less random connections with the image raster. Uhr and Vosslerll performed an important pioneering study of a program that "generates, evaluates, and adjusts" its own parameters. Not surprisingly, however, the automatically generated features were confined to simple, local templates. The recognition of characters printed subject to |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://www.computer.org/csdl/proceedings/afips/1968/5072/00/50721125.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
