Research On The Extraction Method Of Geographic Elements And The Recognition Algorithm Of Point Symbols In Color Topographic Maps

The automatic extraction of the geographic elements and the recognition of point symbols are the techniques of image processing and recognition, which relate to computer vision, physiology, computer science and technology and other various subjects. Corresponding theories and practices have laid a good foundation for further researches of these subjects for years. However, there are still many problems to be solved.In this thesis, we first analyze the significance and development of this project, and then point out the main problems. Then according to the basic principles of target recognition by human visual system(HVS), the new ideas are proposed in terms of image preprocessing, image segmentation, linear feature separation, symbol recognition, character grouping and orientation. Finally, a number of experiments have been conducted to varify the algorithms proposed in this thesis. All of these above lay a good foundation for further research in these subjects.The main contributions in this thesis include:1. Image preprocessing and image segmentation for color topographic maps. Aiming at the additive noise in topographic maps and the problems existing in available denoising methods, wiener filtering is used to remove the noise in the shearlet domain, and then an image denoising algorithm via wiener filtering in shearlet domain is proposed. Not only can this algorithm remove the noise, but it can also preserve more detail information. In terms of edge detection, Edges extracted by the methods based on the wavelet transform are discontinuous, and some weak edges can not be detected, due to the fact that the wavelets are limited to directions. Moreover an appropriate edge detection method is proposed based on the multi-direction shear transform. In this method, the shear transform is used to overcome the directional limitation of the wavelets. Aiming at the problems of large amount of computation, high time complexity and inaccurate classification existed in the traditional image segmentation algorithms, a novel image segmentation algorithm, which is based on randomized sampling and multilevel image fusion, is proposed in this thesis. In this algorithm, the large topographic map is randomly sampled first, reducing the amount of data. Then the optimal clustering centers are acquired by fuzzy C-means clustering. Finally, multilevel image fusion is used to fuse the segmented images into the final segmentation maps. This algorithm can increase the efficiency of the image segmentation algorithms, and can improve the effect of the result images.2. Linear feature separation. There exist lots of linear features passing through the background in topographic maps, the colors of different geographic features are close to each other, and there are lots of aliasing and false colors. All these facts pose challenges to linear feature separation. In order to solve these problems, two algorithms are proposed for linear feature separation. The first one is based on the energy density and the shear transform. The proposition of the energy density breaks through the bondage of the color information on which the traditional image segmetation algorithms depend, and the shear transform is used to increase the directions of linear features. The second one is based on the idea of the seed spreading. The seeds can find other pixels as their brothers to connect into groups according to the similarity of the color and energy. Further, these groups are judged as linear features by comparing with the areas around them in the terms of the color and energy. This algorithm can deal with the edge pixels in linear features properly. In addition, aiming at the problems of the lines overlapped by the sand symbols in desert maps, a new algorithm for linear feature separation is proposed. The morphology-based spur removing method is applied to remove the sand symbols after the process of image binarization and morphology-based thinning. Moreover, multi-level spur removing method is used to remove the complex sand symbols.3. The recognition of point symbols. There exist adhesion, cross and overlap between lines and point symbols in topographic maps, so it is difficult to separate point symbols accurately, which brings challenges for recognizing point symbols by traditional methods based on the pattern of extracting before recognition. In this thesis, we propose a novel algorithm to recognize point symbols directly on topographic maps, which combines the shear transform and Line Segment Generalized Hough Transform(LS-GHT). On the one hand, LS-GHT is proposed to represent color and shape features of point symbols more completely, and it is quite insensitive to noise and a certain degree of deformation. On the other hand, the shear transform is introduced to increase the directional features of point symbols. This new method detects point symbols directly, and has a higher recognition accuracy.4. Character grouping and orientation. Characters, which are extracted from topographic maps, should be grouped into text strings before recognition. According to the features of the characters in topographic maps, two methods are proposed. The first character grouping method is present based on the color, size and direction consistency constraints. This algorithm is implemented by background pixel spreading, and can handle multi-oriented, curved and straight text lines of multi-sized characters. Moreover the method based on the graph model is proposed for character grouping, the construction and the simplification of the undirected graph are used to group characters with significant wide spacing into text strings correctly. In addition, the grouped text strings have lots of directions, and a new character orientation method based on connected domain corrosion is proposed. Firstly, the connected components of texts are rotated in many directions; then these connected components in each direction are eroded; finally, the angles of the texts can be identified acoording to their areas after erosion. This method can detect their orientations and rotate the slant text strings to the horizontal direction, thus improving the recognition accuracy of characters.In this thesis, all the proposed algorithms have be verified by lots of test experiments in color topographic maps, and the experimental results show that all these algorithms achieve sound effects. Besides, some algorithms have been applied on the Free Map K9 platform of a research institution in Xi′an, which include the image segmentation algorithm for large color topographic maps based on randomized sampling and multilevel image fusion, linear feature separation based on the energy density and the shear transform, the algorithm for linear feature separation from the desert maps, the algorithm for point symbol recognition based on the shear transform and LS-GHT, character grouping based on the color, size and direction consistency constraints and character orientation based on connected domain corrosion, these algorithms have been applied in the process of the extraction of geographical elements and the reconition of point symbols and characters.