Zhongliang Fu
Algorithm for Fast Detection and Identification of Characters in Gray-level Images
Zhongliang Fu FulingBian Songtao Zhou Qingwu Hu
Wuhan Technical University of Surveying & Mapping,
School of Informatics, Wuhan, 430079, China
Fuzl Q public. wuhan.cneb.com
KEY WORDS: Algorithm, Image processing, Neural networks, Segmentation, Character identification
ABSTRACT
This paper discusses methods for character extraction based on statistical and structural features of gray levels,
and proposes a dynamic local contrast accommodating line width. Precision locating of character groups is
realized by exploiting horizontal projection and character arrangements of binary images along horizontal and
vertical directions respectively. Also discussed is the method for segmentation of characters in binary images,
which is based on projection taking into account stroke width and character sizes. A new method for character
identification is explored, which is based on compound neural networks. A complex neural network consists of
two sub-nets, with the first sub-net performing self-induction of patterns via 2-dimentional local-connected 3-
order networks, the second sub-set linking up a locally connected BP networks performing classification.
Reinforced reliability of the network recognition by introducing conditions for identification denial.
Experiments confirm that the proposed methods possess impressive robustness, rapid processing and high
accuracy of identification.
1. INTRODUCTION
Automatic identification of freight car plays an important role in intelligent management of railway
transportation. It provides an important evidence for attempering vehicle and freightage. At present, automatic
recording system called electronic railway weighing apparatus is adopted in measuring the load of freight car.
But the system can only measures gross weight of freight car, not automatically record its number and
deadweight. This is operated with manual observation and record. Because freight car runs in high speed, some
errors always exist. Therefore, developing automatic identification system of freight car number is valuable.
A feasible way for realization of such an automatic process is real-time acquisition of car numbers image via
CCD cameras controlled by infrared sensors, followed by fast detection and identification of characters using
computer. Key techniques include fast search, locating, segmentation and identification of characters from
image. At last years, some researches about this subject have been done and some systems have been developed.
But existing methods and systems are not robust and real-time, and identification rate is not high.
In fact, because images are acquired out of doors, different condition, for example different time, shadow,
reflection, strong light in background, stain in bodywork and characters, etc, bring automatic detection and
identification difficulty.
Character search includes both characters extraction and locating. Existing algorithms for character extraction
have two kinds: a) static threshold method. Although its process speed is fast, but the adaptability to weather
and illumination condition is weak. b) line detection method. The kind of method includes line detection
operator, Hough transform, contour tracking, line following, etc. All of these operators can't detect wide line
and is sensitive to noise.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 305