INTEGRATED TECHNIQUE FOR AUTOMATED DIGITIZATION OF RASTER MAPS
Serguei Levachkine and Evgueni Polchkov

(continue...)

2 Raster Map Pre-Processing

The main goal of pre-processing is to prepare raster cartographic images in such a way as to simplify them and increase the reliability of their recognition in the automatic system.

Table 3 presents the proposed sequence of operations for the preparation of raster maps for automated recognition with expert judgement on their possible automation, represented using the notation given in Table 2.

Table 2. The degree of automation of operations.

Table 3. The degree of possible automation of pre-processing operations.

We make the following comments.

Restoration. Paper maps subject to storage and usage show signs of wear on their surfaces; spots, scratches, and even notations. Even maps drawn on more stable media, such as cardboard, Mylar, organic glass, or aluminum are subject to the same defects. These signs of wear should be eliminated wherever possible, otherwise they impede the automated recognition of cartographic features.

Copying. If the creator of a GIS project has copying equipment and the GIS is "analytical" not "register", then both manual and automated vectorization are usually performed on copies of the map rather than on the original. The advantages include preservation of the original, parallel processing of the vectorization by several manual operators and registration of the graphical singularities of the objects of a given class for further processing of the copy.

Increasing the contrast of object images. In order to simplify the process of vectorization, objects of the same class can be highlighted on a copy of the map using a contrasting color. Typically such marking involves enclosing linear objects such as rivers, roads or pipelines. In practice, outlines of polygonal objects which do not have explicit borders (such as bogs, bushes, etc.), and are delineated only by dashed or patterned lines, must be drawn in. In particular, various polygonal objects may overlap, one represented by color, another outlined by a dashed line, and a third by patterned lines; in such cases, the objects must all be outlined explicitly.

Choice of scanning parameters. Both existing research and our experience in image recognition show that the efficiency of object identification depends on the choice of scanning parameters. Candidates are resolution, color palette, contrast, brightness and visual effects. It is known that an optimal set of scanning parameters exists both for every scanner and for each map to be scanned ([7]). The parameter set can only be tested by repeated test scans followed by visual or automated analysis of the raster images obtained. An objective justification of the choice of optimal scanning parameters can be obtained by statistical calculation of the raster characteristics. Criteria for estimating the statistics can be developed from the point of view of effectiveness of the cartographic image recognition.

Joining raster facets. The necessity of joining raster facets arises for two main reasons; first, large-scale scanners are very expensive, and most users do not have access to them, and second, the cartographic materials required for the production of GIS often have different standards for different scales. In the present work, we may assume that software is available which can edge-match raster facets with minimal error ([1],[7]).

Correction of raster image geometry. In practice, the raster image obtained after scanning is not uniformly deformed, due to printing errors, wear, scanning errors, and defects in edge-matching. Raster transformation programs exist to eliminate or at least minimize these defects. This has the direct effect of increasing the accuracy of the vectorization of the final map and the indirect effect of ultimately improving image recognition. The general principle of raster map correction is that of plane transformation by reference points, i.e. displacement of certain points of the raster map whose coordinates are known; followed by translation of the remaining elements of the raster correspondingly. Reliablity of raster map correction is maximized when geodetic reference points are used as the control points. A satisfactory correction to the raster map can be provided by a coordinate grid on the original map. In this case, if the type and parameters of the cartographic projection are known, programs can be developed which generate a theoretically exact system of the reference points used for transformation of the raster map. If neither geodetic data nor a coordinate grid are available, larger-scale or same-scale vector maps which have already been corrected, or satellite images containing the reference points can successfully be used to correct the raster map. In this case, point features of the cartographic objects, such as river confluences, road intersections, bridges, etc. can be chosen as control points.

Elimination of map notations and legend. After scanning, a raster map contains much graphical information that it is not necessary to vector, such as toponyms, annotations, explanations, copyright notice, etc. This information must be eliminated to simplify cartographic image recognition. The map legend must also be eliminated, and processed as a separate raster image.

Elimination of artifacts. Artifacts (traces of folds, stains, scratches, dust, etc.) which contain no information are nearly always present on a raster map, and complicate subsequent cartographic object recognition. A difficulty in processing artifacts is the necessity of restoring the image previously obscured by the artifact.

Restoration of the topology of cartographic images. As a result of printing and scanning errors, graphical images of the cartographic objects on a raster map frequently have topological defects. Typical defects are breaks in thin lines which should be continuous (such as contour lines, rivers, etc.) and the fusion of lines that logically should not intersect (eg. contour lines). Topological errors in raster images complicate the recognition of cartograpic objects, and gross errors which can be noted by visual analysis of a raster map must be corrected before the automated recognition procedure is begun. The difficulty of solving this problem is increased by the fact that such defects are often only visible, and can only be corrected, at the pixel level. Nevertheless, powerful programs for correction of raster maps at pixel level currently exist, providing hope for the solution of this problem ([1], [3-13]).

Separation of basic colors. In graphical coding systems for cartographic information, a limited number of colors is used for a given map. However, after scanning, a raster map contains a wide palette of colors, complicating image recognition. Analysis of the map legend colors and selection of the best quality segments in the raster field aid in reconstructing the basic graphical coding palette. Superfluous colours and their shades can then be substituted by the corresponding elements of the basic palette.

Stratification of raster map. A raster map, considered as a unified heterogeneous graphical image, is suitable for parallel human vision. In contrast, raster images, containing homogeneous graphical information, are suited to consecutive machine vision. Two approaches can be used for the stratification of the original raster map; stratification by reduced color palette, or logical stratification of the cartographic objects.

In the first case, maps are derived from the original raster map which preserve only those pixels which have a strongly defined set of colors corresponding to the images of one class (for example, red and yellow, say, might correspond to the icons of populated places and the color of their outlines). In the second case, the map only preserves fragments of general raster images corresponding to the locations of cartographic objects of one class (for example, railways with adjacent thin zones).

The procedure of stratification by color is clear, therefore let us consider here the basic features of logical stratification of a raster map. The principle is that the presence of a cartographic object, which must be vectored in logically separated layers, is reliably known. Thus, the only task left for the recognition program is specification of the location of the object. This simplifies the problem and increases the reliability of the solution. Logical stratfication of a raster map can be done by hand. An operator moves a window of given size and shape, separating the fragments of the layer formed. The efficiency of logical stratification, even when performed manually, lies in that the main difficulty in vectoring raster maps is visual fixing of the coordinates of an object. In practice, this means that the operator must change the scale of the image and fix its coordinates using the cursor, perhaps two or three times, until a satisfactory result is obtained. On the other hand, defining the location of the object by specifying the size of a window is easier, especially when it is taken into consideration that the operator need not be concerned with any overlap of raster map facets, since the program can correct for this.

A typical situation that may arise is that a small scale map already exists for a given territory. In this case, for logical stratification of a raster map, one must use methods for constructing buffer zones of the linear objects. These are already available in some popular vector editors ([2], [7] and [16]).

Dirección General de Servicios de Cómputo Académico-UNAM
Ciudad Universitaria, México D.F.