IEEE TRANSACTIONS ON SYSTEMS , MAN , AND CYBERNETICS

These results indicate that shadow detection is difficult prior to segment noise cleaning, but may be possible for the nonnoise segments. Shadows can be used to verify the recognition of segments belonging to buildings and roads; the former should have shadows if they are appropriately oriented, but the latter should not. VII. CONCLUDING REMARKS The approach used here is quite straightforward. It proceeds in an essentially bottom-up fashion, with no provision for feedback between levels, and it makes no use of higher-level information, e.g., that buildings are alongside roads, or that roads form a connected network. Its relative success in spite of these restrictions illustrates the possibility of achieving reasonable performance with a simple bottom-up approach. The programs described in this correspondence made use of a number of empirically chosen constants. In some cases, these simply represented liberal thresholds defined by round numbers (e.g., 9, 0.2, 8, 5 percent, and the rate of linear falloff in Section Ilff; two-thirds and 1550 in Section III-A; 25° and 10 percent in Section III-B; etc.) In other cases, they were based on information about scale (i.e., the sizes (in pixels) of the buildings and roads that were to be detected; cf. the four-pixel strip width used in Section II) or grayscale (i.e., their contrasts; cf. the ten gray level range in Section Ill-A), and would have to be adjusted for different types of imagery. In any case these parameters worked well for all five of the examples on which the program was tested [8], two of which are given here, as well as for two other examples taken from an aerial photograph of a different part of the country. These programs were not designed to be computationally efficient ; their running time was 10-20 min on a time-shared Univac 1108. It is evident, however, that the method used here could be implemented very efficiently using suitable parallel hardware, since the processing of segments is largely local. Many specific improvements in the approach are possible at each of its stages. The process of fitting straight line segments to connected components of edge pixels is somewhat order-dependent , and tends to produce overshoots; it might be better to use a Hough-like approach to detect clusters of collinear edge pixels. Rather than making a succession of decisions about segments, pairs of segments, and groups of segments, it might be better to design a hierarchical relaxation scheme …