Map Character Extraction And Recognition Algorithms

The automatic input and recognition of topographic maps is widely recognized as one of the most difficult problems in the world. It is the key technique to manage and publish map information automatically. The managing technology of database in geographic information system has become relatively mature, and it becomes the bottleneck of the development for the system to obtain data. There are many kinds of symbols in the topographic map. The contour, point symbol and Chinese lettering annotation have been chosen as the main object of research in this thesis, and the valid method will be extended to other symbols, thus all the symbols can be extracted by the similar way. There are also different topographic map scales. As the information density in the topographic map of 1:50000 is the biggest, and other topographic map of several kinds of scales can be obtained from it, so the scale of 1:50000 is chosen as the scale of our research map.At the very beginning of this dissertation, the development of extracting and recognizing symbols of the map is summarized. In the following content, the algorithms related to the thesis are simply introduced.On the aspect of color layer, a 4-D model, which combines the pixel's planar distribution with the color space organically, has been structured on the base of the human eye's mechanism, the color science and the imitation of the function of the human eye on color vision.The synthetic algorithm of color gathering based on the 4-D model has been proposed. A linear transformation of the 4-D space has been taken and different characteristic functions have been structured according to four kinds of topographic elements separately. Aiming at the color error produced in the process of print and scan, the algorithm has separated the layers of different topographic elements by synthesizing information and decomposing components in layer. A large number of tests shows that the algorithm is efficient, widely adapted and highly automatic.On the aspect of the recognition of contour, the idea that the thick contour should be separated from the thin contour before recognizing the whole contour is offered. It is to say that the general methods of pattern recognition are used initially, recognition is applied in different layers, objects are roughly categorized first, then they are processed more, but the useful information have been preserved to guide the later recognition during the processing. The new idea provides the significant pre-knowledge for the connection of the contour, so that the connecting rate of the disjunctive contour has been greatly raised.In the connecting of the contour, a multi-layer hybrid method has been proposed. A pyramid structure is established, and it makes the distribution of human intelligence and computer intelligence the most reasonable. A knowledge base has been set up for the broken points of contour. The knowledge base, which combines the dynamic structure with static ones, increases the efficiency of program obviously.The concept of contour groups and mass contours has been given. When the contour dataare compressed into vector, analyzed by topology and given the elevation value, the contour data has been structured reasonably according to the knowledge of topographic map. An algorithm, which calculates the contour elevation automatically based on bin tree, is presented too. A high speed and accuracy can be acquired in this method.As to the point symbol, an algorithm based on the information fusing of basic elements and the analysis of confidence has been adopted; the algorithm uses the multi-character as the foundation of recognition. Two kinds of neural network classifiers are integrated in it. The final result is synthetically judged by the confidence in the classifier's output data, and at the same time the flow is controlled by the threshold of confidence, so the total system will reduce the cost of time without any decline in recognition rate.On the aspect of the extraction of Chinese lettering annotation, a recognition framework is built by the related t...