MARIS: Map Recognizing Input System - Implementation and Performance on a Special Workstation

1. I n t r o d u c t i o n Interactive digitizers are used as input units in conventional geographical information systems because of their high d a t a compression rate, low hardware cost and operational simplicity [I]. However, considerable time and cost are necessary to digitize geographical da ta because of the number of manual operations involved in tracing. These manual operations can be reduced by scanner systems which are suitable for the automatic digitization of maps or aerial photographs. Digitization should produce information in a layered da ta form, because this is needed for easy display, intelligent retrieval and efficient storage [2]. However, this has been considered difficult and different input techniques are essential to overcome this problems. We have already proposed an automatic map recognizing input system called MARIS [3]. This system digitizes large-scale maps into a layered data form in order to construct a map database. This paper presents the implementation of the MARIS system and assesses its performance on large-scale national maps of Japan. The MARIS system is implemented on an experimental workstation which consists of a main control unit and a graphic processing unit. The main control unit is the same type of workstation as used for distributed processing of multi-media da ta 141. The graphic processing unit has multi-processors, largecapacity window map memory, magnetic disk storage for enhancement of large-scale image processing speed. The MARIS process consists of four stages: preprocessing, vectorization, automatic recognition, and MARIS's interactive input times are shorter than those using conventional interaction methods, because the number of manual operations are reduced by utilizing basic line tracking algorithms. 2. E x p e r i m e n t a l W o r k s t a t i o n The MARIS is implemented on an experimental workstation system. The system consists of a scanner, a workstation (called WS-R), magnetic tape storage, and an electrostatic plotter as schematically shown in Fig. 1. Equipment specifications are shown in Table 1. The experimental workstation consists of two units: the main control unit and the graphic processing unit. The main control unit operates under UNIX System V. This unit is the same type of workstation as used for distributed processing of multimedia d a t a [4]. The main control unit is composed of a Motorola 68020 processor, an 8 Mega-byte main memory, 320 Mega-byte magnetic hard disk storage, and a local area network controller. The graphic processing unit consists of a graphic engine, 32 Mega-byte window map memory, a CRT, frame memory, 110 interfaces t o a large-scale scanner and a plotter, and 80 Mega-byte magnetic disk storage for image data. The graphic engine is composed of a Motorola 68020 processor and four data flow pipeline processors [5]. T h e ~ i ~ e l i n e processors are ~rogramrnable and enhance the speed of image processing operations. If an image processing operation is simple and is suitable for parallel processing, pipeline processors enable high-speed image processing. Otherwise, the Motorola 68020 processor is used. The MARIS system performs thinning operation, feature point extraction, rotation, etc., by the pipeline processors. T h e graphic processing unit has large capacity window map memory. The capacity is enough to IAPR Workshop on CV Speaal Harchare and Industrial Applications OCT.12-14. 1988. Tokyo store a complete digital binary image of an input map sheet. Because this memory is used as working memory in vectorization and interactive correction, the image da ta is stored on the 80 Mega-byte magnetic disk. This capacity is sufficient for two map sheets. 3. I m p l e m e n t a t i o n o f MARIS T h e MARIS process consists of the four stages : preprocessing, vectorization, automatic recognition, and interactive correction. Preprocessing and vectorization are performed on the graphic processing unit. This unit also carries out image input processing and image output processing including display processing. Most of the automatic recognition and interactive correction are performed by the main control unit. 3.1 P r e p r o c e s s i n g A large flat scanner initially converts a map sheet into a binary image directly by grid sampling a t a resolution of 16 pixels per mm. After input, the binary image is rotated 90 degrees, and stored in the window map memory. T h e pipeline processors directly contribute t o increase in the rotation speed. The picture is divided into 70 subpictures of 2048 pixels x 1024 pixels and stored on the magnetic disks. These subpictures are processed one by one as described below. 3.2 Vec tor iza t ion The vectorization stage consists of labeling by line widths, thinning, transformation of an image into a graph, edge deletion, and straight line approximation. First, a subpicture is transferred from the magnetic disk into the window map memory. Then, all 1-pixels in the subpicture are labeled according to the width of the lines passing through them [6]. Next, the labeled subpicture is converted into an 8-connected medial line image by thinning operation [7]. Then, the medial line image is transformed into a graph, in which pixels on the medial line correspond t o nodes and neighboring nodes are connected by edges. Next, extra edges, unnecessary for preserving the topology of the medial line image, are deleted to form a simplified graph [a]. Next, vector data is produced from the simplified graph by extracting feature points and approximating curved lines between two feature points as straight line segments. The vector data is transferred to the main memory of the main control unit. Fig. 2 shows a n example of the vectorization process. The pipeline processors enhance the speed of the thinning operation and the feature extraction. For example, a thinning operation using the pipeline processors runs 20 times faster than when using a general-purpose VAX 11/780 computer. Width labeling, transformation of a n image into a graph and edge deletion could be performed in parallel, but instead they are performed on the graphic engine CPU. This is because the program area of the pipeline processors is too small to store their programs. Extracted vector da ta are composed of feature points, branches and segments as shown in Fig. 3. A curve between the two feature points is called a branch. Line segments produced by a straight line approximation of a branch are referred to as simply segments.