Thesis for scientific degree of Candidate of Technical Sciences in specialty
05.13.06 - Information technologies. - National Technical University "Dnipro Polytechnic", Dnipro, 2021.
This dissertation work provides a solution for developing information technology building recognition on images from remote Earth sensing and verification of recognition results. Dissertation work contains general information about photogrammetry, current state of Earth remote sensing, general approaches of pattern recognition, including existing software, methods and technologies for image processing with high spatial resolution and there are currently several information technologies for automatic building recognition showing low accuracy. The developed technology consists of the following stages: division of a scene into sites; histogram analysis; feature segmentation; verification of recognition results based on expert database; shadow analysis; geometry analysis; building contour localization.
In the first stage, the image is divided into segments to localize the search and to make the initial simplification of the recognition scene and to determine plot types (residential, multi-storey or commercial), which facilitate the recognition process. The second stage - histogram analysis is based on the localization of peaks in the histogram. In a large sample of buildings in the test images, it was observed that about 50% of the plots contain buildings that create a majority peak (the highest peak on the histogram). In other cases, the building generates a peak, but not a majority, and in some cases,
there is more than one house or complex of buildings on the site. Analysis of histograms makes it possible to determine the binarization threshold at the next stage - the segmentation stage. The building contour is removed from the binary image based on the Suzuki-Abe algorithm.
The following steps are a mechanism for estimating the probability of a segment being part of a building. Size analysis is based on expert attribute data. The size constraint can be used to eliminate segments with features that do not match the expert data about the site. The size constraint is estimated either by the area taken from the attribute table or calculated as the minimum percentage of the plot size. After weeding out small segments, the neural network, previously trained on expert data, decides to eliminate the "non-building" segment. At the stage of shadow analysis, belong to the segments with buildings, and segments with "incorrect" shadows, and which are not buildings. Pixels with certain values are grouped into segments ("feature" segments and shadow segments). Since the shadow segment and the building segment should be adjacent, a buffer is created around the segmented shadows. Each shadow segment is then examined for possible overlaps with buffers for "special" segments as there might be more than one shadow area around the building. Any segment that is overlapped by the shadow buffer is marked as a potential building. If the shadow is located on the "wrong" side, the segment is removed.
Elimination of segments with a low probability of being a building also depends on geometry. Measures used for geometric analysis were selected as follows – rectangular, round, monolithic, convex. These characteristics are checked individually by comparing the behavior of the parameter for objects "building" and "non-building". The values of each parameter were used to calculate the probability of the segment "being a building". The possibility of cavities within a segment is assessed as an indication to exclude a segment. For example, cavities larger than expected will be an indicator of the absence of a building. Simple restrictions are used, such as the minimum width of the building. Segments with features that are defined on the basis of various measures not as buildings will be eliminated. At the last stage, using the Ramer- Douglas-Packer algorithm, the obtained contour of the building is smoothed, and the raster image is converted into a vector one.
Proposed technology was tested with three different characteristic types of scenes: multi-storey buildings, commercial (industrial) buildings and residential single- home areas. The evaluation of obtained results was performed by comparing the area and geometry of the removed image and the parameters of the test building. The test results showed that commercial (industrial) buildings are the most recognizable. Trees, shadows and the offset for the terrain interfere with the building recognition.
The proposed information technology is implemented in the form of a software application, and it allows performing operations on building recognition from any photogrammetric images of various nature (aerospace, lidar, quadrocopters) even by inexperienced users and significantly reduces the time for obtaining results and improves the quality of recognition.
