Automatic input of line drawings with contour-based picture processing, vectorization and loop recognition

Abstract

Automatic input of paper-based line drawings into a CAD system can be performed by converting the rasterized image of the drawing into a description in terms of CAD primitives. The conversion process can be roughly divided into two functional modules of low-level picture processing and application specific meaning extraction. The automatic input system should be accurate, robust, and efficient. Most of existing systems impose severe restrictions on the input drawings to increase their recognition accuracy, but it prevents them from being totally effective in the general engineering environment. In addition, little care has been taken about the computational efficiency. In this thesis, we present two automatic systems for the recognition of logic circuit diagrams and the vectorization of general line drawings. By constructing the preprocessing module which is commonly used for the two systems as a sequence of contour-based picture processing operations, we can make them highly efficient. Through and error correcting mechanism and a flexible knowledge representation, we can make them reliable for the syntactic variations of input drawings. In case of vectorization, we can obtain higher degrees of data compaction ratio and applicability from the extraction of free-form curves. Experiment with a number of logic diagrams shows that the recognition system identifies more than 96＼% of symbols and connections and that an A4-sized diagram with an average complexity can be processed within 15 seconds on a workstation. Another advantage of the system is the flexibility of knowledge representation in which the prior knowledge about input diagrams can be modified without rewriting the recognition program. Similar performance can also be obtained for the vectorization system. The efficiency in terms of time and storage can be readily achieved by the contour-based picture processing, and higher degrees of geometric accuracy and data compression can be reached from the v...