JP2013012034A - Area extraction method, area extraction program and area extraction device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique for extracting an area of a building more correctly, on the basis of data of an aerial photograph or the like.SOLUTION: When an area of a building is extracted on the basis of data of an aerial photograph or the like, multiple different discretization widths are set, and for each of the discretization widths, a luminance value of the data is discretized to multiple values discretized and set by the discretization width. In a discretized image obtained by the discretization, pixels having the same value are connected, and an area having a shape close to a rectangle is extracted as a candidate for the area of the building. Subsequently, among the multiple areas extracted for the respective multiple different discretization widths, an area having the shape closer to the rectangle is adopted as the area of the building.

Description

  The present invention relates to a method, a program, and an apparatus for extracting a building area based on photograph data taken from an aircraft or an artificial satellite.

  Feature extraction and area extraction in the case of bird's-eye view have been important themes from the past (for example, Non-Patent Document 1). For example, road extraction from aerial photographs or artificial satellite images has been actively studied (for example, Non-Patent Documents 2 and 3). Remote sensing image classification methods can be broadly classified into pixel-based classification and area (object) -based classification. Pixel-based classification techniques such as clustering (so-called unsupervised), maximum likelihood (supervised), and Support Vector Machines (SVM) (supervised) assign each pixel to a classification class based on statistical theory. (For example, Non-Patent Document 4). On the other hand, in the region-by-region classification method, information around the pixel (Context) is used. For example, classification using a mathematical Morphological approach is a typical example (for example, Non-Patent Documents 5 to 9).

  (A) and (b) of FIG. 1 are examples of aerial photographs of an urban area (Higashiyama Ward, Kyoto City) where houses are densely packed. In such a dense urban area, the conventional region extraction method does not function effectively. When analyzing the factors, first, (a) because the buildings are close to each other, if the brightness value and texture of the roof are similar, the boundary line of the building will not be clear. In addition, (b) there is a factor peculiar to dense urban areas that the shadows of adjacent buildings are easily reflected. Furthermore, roofs with corrugated cross-sectional shapes (for example, slate roofs) are sometimes used in traditional Japanese buildings, and (c) it is difficult to extract the contours of roofs with textures that vary in brightness. The factor due to the building characteristics in the target area can also be confirmed.

Weng, Q. and Quattrochi, D.A., "Urban Remote Sensing", CRC Press, 2007 Hu, J., Razdan, A., Femiani, JC, Cui, M. and Wonka, P., "Road network extraction and intersection detection from aerial images by tracking road footprints", IEEE Transactions on Geoscience and Remote Sensing, Vol. 45, No. 12, pp. 4144-4157, 2007 Movaghati, S., Moghaddamjoo, A. and Tavakoli, A., "Road extraction from satellite images using particle filtering and extended Kalman filtering", IEEE Transactions on Geoscience and Remote Sensing, Vol. 48, No. 7, pp. 2807- 2817, 2010 Tso, B. and Mather, P.M., "Classification methods for remotely sensed data-2nd ed.", CRC Press, 2009 Soille, P. and Pesaresi, M., "Advances in mathematical morphology applied to geoscience and remote sensing", IEEE Transactions on Geoscience and Remote Sensing, Vol. 40, No. 9, pp. 2042-2055, 2002 Benediktsson, JA, Pesaresi, M. and Amason, K., "Classification and feature extraction for remote sensing images from urban areas based on morphological transformations", IEEE Transactions on Geoscience and Remote Sensing, Vol. 41, No. 9, pp. 1940-1949, 2003 Benediktsson, J.A., Palmason, J.A. and Sveinsson, J.R., "Classification of hyperspectral data from urban areas based on extended morphological profiles", IEEE Transactions on Geoscience and Remote Sensing, Vol. 43, No. 3, pp. 480-491, 2005 Bellens, R., Gautama, S., Martinez-Fonte, L., Philips, W., Chan, JC-W., Canters, F., "Improved Classification of VHR Images of Urban Areas Using Directional Morphological Profiles", IEEE Transactions on Geoscience and Remote Sensing, Vol. 46, No. 10, pp. 2803-2813, 2008 Tuia, D., Pacifici, F., Kanevski, M. and Emery, WJ, "Classification of Very High Spatial Resolution Imagery Using Mathematical Morphology and Support Vector Machines", IEEE Transactions on Geoscience and Remote Sensing, Vol. 47, No. 11, pp. 3866-3879, 2009 Canny, J., "A Computational Approach to Edge Detection", IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. PAMI-8, No. 6, pp. 679-698, 1986 Zhong, J. and Ning, R., "Image denoising based on wavelets and multifractals for singularity detection", IEEE Transactions on Image Processing, Vol. 14, No. 10, pp. 1435-1447, 2005 Tonazzini, A., Bedini, L. and Salerno, E., "A Markov model for blind image separation by a mean-field EM algorithm", IEEE Transactions on Image Processing, Vol. 15, No. 2, pp. 473- 482, 2006 Berlemont, S. and Olivo-Marin, J.-C., "Combining Local Filtering and Multiscale Analysis for Edge, Ridge, and Curvilinear Objects Detection", IEEE Transactions on Image Processing, Vol. 19, No. 1, pp. 74 -84, 2010 Ding, J., Ma, R. and Chen, S., "A Scale-Based Connected Coherence Tree Algorithm for Image Segmentation", IEEE Transactions on Image Processing, Vol. 17, No. 2, pp. 204-216, 2008 Chien, S .Y., Ma, SY and Chen, LG, "Efficient moving object segmentation algorithm using background registration technique", IEEE Transactions on Circuits and Systems for Video Technology, Vol. 12, No. 7, pp. 577-586 , 2002 Tsai, V.J.D., "A comparative study on shadow compensation of color aerial images in invariant color models", IEEE Transactions on Geoscience and Remote Sensing, Vol. 44, No. 6, pp. 1661-1671, 2006 Ma, H., Qin, Q. and Shen, X., "Shadow Segmentation and Compensation in High Resolution Satellite Images", Proceedings of IEEE International Geoscience and Remote Sensing Symposium, 2008, Vol. 2, pp. II-1036-II -1039, 2008

  In other words, the above factor (a) gives a problem of how to deal with the case where the edge of the region is not sufficiently obtained. It is also a problem to exclude unnecessary edges from edges that are excessively extracted due to the above factor (c). Here, it is considered that the problem (c) is related to the problem (b). For example, in Non-Patent Document 10, Canny has proposed an edge extraction operator that is resistant to noise. This is one of the typical edge extraction operators widely used so far. In addition, a method of removing noise using Wavelet (for example, Non-Patent Document 11) or utilizing an average value (for example, Non-Patent Document 12) or a multi-resolution image (for example, Non-Patent Document 13) has been reported. ing.

  In connection with the above factor (b), research cases for restoring the luminance value of the shadow region have been reported (for example, Non-Patent Documents 14 to 17). However, for example, as far as the restoration result in Non-Patent Document 17 is seen, there is a problem that the boundary line between the original shadow area and the non-shadow area appears clearly in the restored image, and the correction value in the shadow area Problems such as difficult decisions and sometimes overrecovering are still unsolved.

  In view of such conventional problems, an object of the present invention is to provide a technique for more accurately extracting a building area based on data such as aerial photographs.

  (1) The present invention is an area extraction method for extracting an area of a building based on data of a photograph taken from an aircraft or an artificial satellite, and sets a plurality of different discretization widths. The luminance value of the data is discretized into a plurality of values set discretely with the discretization width, and pixels having the same value are connected in a discretized image obtained by discretization, and is close to a rectangle as a candidate for a building region A region having a shape is extracted, and a region having a shape closer to a rectangle is adopted as a region of a building among a plurality of region groups extracted for each of the plurality of different discretization widths.

  In the region extraction method as described above, a region having a texture with a large dispersion of luminance values can be extracted as one region by discretization. Further, by connecting points having the same value and extracting a region having a shape close to a rectangle, it becomes easy to remove regions other than buildings. In addition, by adopting a region of a more rectangular shape as a building region among a plurality of region groups extracted for a plurality of different discretization widths, a locally optimal spatial smoothing parameter A function equivalent to that of applying can be exhibited.

(2) In the region extraction method of (1), a rectangular index defined by (area of area / area of rectangle surrounding area) can be used as an index representing a shape close to a rectangle.
In this case, the degree of “closeness” in a shape close to a rectangle can be expressed simply and accurately as an index.

(3) Also, in the region extraction method of (1) or (2), when the adjacent regions are merged in the extraction, the merger can be executed if the region becomes closer to a rectangle. It may be.
In this case, the area of the building can be captured more accurately.

(4) In the area extraction method of (2) above, when the rectangular index is smaller than a predetermined value, it is preferably not adopted as the area of the building.
In this case, an area that is highly likely not to be a building can be excluded from extraction targets.

(5) In the region extraction method of (1) above, a region estimated to be vegetation may be excluded from the extraction target based on the RGB luminance values.
In this case, the region can be excluded from the extraction target in consideration of the color characteristics of the vegetation.

  (6) On the other hand, the present invention is an area extraction program for extracting a building area based on data of a photograph taken from an aircraft or an artificial satellite, and sets a plurality of different discretization widths. A function of discretizing the luminance value of the data into a plurality of values discretely set with the discretization width, connecting pixels having the same value in the discretized image obtained by discretization, and candidate building areas A function of extracting a region having a shape close to a rectangle, and a function of adopting a region having a shape closer to a rectangle out of a plurality of region groups extracted for each of the plurality of different discretization widths as a region of a building This is an area extraction program to be realized by a computer.

  (7) Further, the present invention is an area extraction device that extracts a building area based on data of a photograph taken from an aircraft or an artificial satellite, and sets a plurality of different discretization widths, and each discretization width is set. For the function of discretizing the luminance value of the data into a plurality of values discretely set with the discretization width, and connecting pixels having the same value in the discretized image obtained by discretization, A function of extracting an area having a shape close to a rectangle as a candidate, and a function of adopting an area having a shape closer to a rectangle out of the plurality of area groups extracted for each of the plurality of different discretization widths, as a building area It is what has.

In the area extraction program / area extraction apparatus of (6) and (7) above, an area having a texture with a large variance of luminance values can be extracted as one area by discretization. Further, by connecting points having the same value and extracting a region having a shape close to a rectangle, it becomes easy to remove regions other than buildings. In addition, by adopting a region of a more rectangular shape as a building region among a plurality of region groups extracted for each of a plurality of different discretization widths, a locally optimal spatial smoothing parameter can be obtained. The function equivalent to what is applied can be exhibited.

  According to the present invention, it is possible to more accurately extract a building area based on photograph data such as aerial photographs.

It is an example of an aerial photograph of an urban area (Higashiyama-ku, Kyoto) where houses are densely packed. This is a photograph taken on the ground around Higashioji Road from Hokanji Temple in Higashiyama-ku, Kyoto. It is a block diagram which shows an example of the computer apparatus for implementing an area | region extraction method. It is a conceptual diagram of the connection of a pixel. It is a conceptual diagram of rectangular index calculation. It is a conceptual diagram of integration of a target region and a region adjacent thereto. It is a figure which shows the result of the edge extraction in the urban area (area 1) where a low-rise building is crowded. It is a figure which shows the result of the edge extraction in the urban area (area 2) where a high-rise building and a low-rise building are mixed. It is a figure which shows the result of the edge extraction in the area (area 3) where Takagi is adjacent to a building. It is a figure which shows the result of the edge extraction in the area (area 4) where many buildings with a dormitory roof exist. It is a figure which shows the result of the area | region extraction in the area 1. FIG. It is a figure which shows the result of the area | region extraction in the area 2. FIG. It is a figure which shows the result of the area | region extraction in the area | region 3. FIG. It is a figure which shows the result of the area | region extraction in the area | region 4. FIG. It is a figure which shows the comparison result of the area | region extraction about the area | region 1. FIG. It is a figure which shows the comparison result of the area | region extraction about the area | region 2. FIG. It is a figure which shows the comparison result of the area | region extraction about the area | region 3. FIG. It is a figure which shows the comparison result of the area | region extraction about the area | region 4. FIG.

  Hereinafter, a region extraction method / region extraction program according to an embodiment of the present invention will be described. As a model area for region extraction, the area around Higashioji Road from Hokanji Temple, Higashiyama-ku, Kyoto was targeted. In this area, wooden buildings are lined up across the entrance. FIG. 2 is a photograph of the area taken on the ground. Further, as the aerial photograph data of this area, data (commercially available product) having been subjected to ortho (orthogonal projection) processing at 25 cm resolution was used. This data is aligned with the planar rectangular coordinate system. In addition, even if it has been ortho-processed, the side of the building can be seen slightly, but there is no particular problem in practical use.

  Next, an area extraction method and area extraction program for extracting a building area (roof outline shape) will be described. FIG. 3 is a block diagram showing an example of a computer apparatus 10 for carrying out the area extraction method, and the area extraction program causes the computer apparatus 10 to realize a function. In addition, the computer apparatus 10 having the function of the area extraction program is an area extraction apparatus according to an embodiment of the present invention.

  The computer 10 includes a CPU 1, a memory 3 connected to the CPU 1 via a bus 2, an auxiliary storage device 4 such as a hard disk, interfaces 5, 6 and drives 7 connected to the interfaces 5 and 6, respectively. And a display 8, typically this is a personal computer. By attaching a storage medium 9 such as a CD or DVD containing aerial photographs to the drive 7 as an input device, the aerial photograph data can be read. The area extraction program is installed and executed in the auxiliary storage device 4 via a storage medium or a network separately from the aerial photograph. That is, the area extraction program can exist and be distributed as a storage (recording) medium.

《Region extraction procedure》
Hereinafter, region extraction (a main step of the method / program) will be described in detail.
As an outline of region extraction, in order to effectively extract a roof having a texture with a large variance of luminance values, the luminance values are discretized into a predetermined number of values, and regions having the same discretized values are labeled. And the area | region of the shape close | similar to the rectangle seen in the building seen planarly is extracted preferentially. In addition, for regions obtained by applying a plurality of different discretization widths, the spatially smoothing parameters that were optimal locally were applied by sequentially adopting regions that were closer to a rectangle. The same processing is realized. Hereinafter, each step will be described in detail.

(First step)
First, N _disc types of discretization widths are set for the luminance values of an arbitrary 1 band of 1 byte (luminance value range: 0 to 255) × 3-band (RGB) image. Under each discretization width, N _off types of different offset values are applied to discretize the image luminance values. For example, next to "offset value width = 8" when the "discretization width = 40 of the luminance value""Offset Number = 5", the _{N off} type of "offset = {0,8,16,24,32}" A discretized image is obtained. Under “offset value = 0”, the luminance value “0-39” of the original image is given the same discretization value, and similarly “40-79” “80-119” “120-159” “160-199”. ”“ 200 to 239 ”and“ 240 to 255 ”. In this experiment, the following parameters were used as an example.

Band used: Red (R) band Discrete width Δd of luminance value = {40, 30, 20}
Number of offsets N _off = 5
Offset value width Δ _off = Δd / N _off = {8, 6, 4}

(Second step)
Next, in each discretized image, four directions (up, down, left and right) are searched, and pixels having the same value are connected and extracted as regions. A large region (for example, 6400 pixels or more) having a certain area or more is removed. In addition, small regions less than a certain area (for example, less than 80 pixels) are integrated if there is a region larger than a certain area in the periphery, and are removed if there are regions. Thereafter, the edge (contour) of each region is extracted.

(Third step)
Next, all the edges obtained from N _off types of discretized images in a certain discretization width are superposed on one sheet.

(4th step)
Next, an edge having a certain strength or higher is left, and an edge connected to the edge is also extracted. At this time, if adjacent roofs and buildings have almost the same brightness value, those edges are often not extracted, so it is confirmed that there are straight edges around even if there are no edges If possible, join the edges.

This edge connection will be supplementarily described. First, eight directions (upward, downward, leftward and rightward) around a pixel that is not an edge are searched, and the number of edges existing in the fixed pixel number d is examined.
FIG. 4 shows an example of searching for the edge group 1a in the upper left direction and the edge group 1b in the lower right direction. For example, (a) group 1a has d1 or more edge groups, and group 1b has d1 or more edge groups, (b) group 2a has d2 or less edge groups, and group 2b has only d2 or less edge groups. If both conditions (a) and (b) are satisfied, the non-edge center pixel is complemented as an edge. The condition (b) prevents the pixels inside the area near the corners of the rectangular area from being erroneously extracted as edges. The search is performed four times, in the vertical direction, left and right direction, upper left to lower right, and upper right to lower left. This time, d = 7, d1 = 5, and d2 = 1 were set.

(5th step)
Next, a label number is assigned (labeled) to the region closed by the edge, and the region that seems to be vegetation is removed based on the RGB luminance value. As in step 2, a large area (for example, 6400 pixels or more) having a certain area or more is removed, and a small area having a certain area or less (for example, 30 pixels or less) is integrated if there is a certain area or more in the periphery. Remove.

A supplementary explanation will be given for the removal of vegetation. In the target area, green and red vegetation were confirmed. for that reason,
(1) B _grn ≧ T _{veg, DN} , B _blue ≧ T _{veg, DN} , and B _grn / B _blue ≧ T _{veg, g2b_ratio}
(2) B _red ≧ T _{veg, DN} , B _blue ≧ T _{veg, DN} , and B _red / B _blue ≧ T _{veg, r2b_ratio}
The ratios R _{grn_veg} and R _{red_veg} occupied by pixels satisfying any of the _above are calculated. Here, B _red , B _grn , and B _blue represent the luminance values of the red, green, and blue bands, T _{veg and DN} represent the luminance value threshold values, and T _{veg, g2b_ratio} , and T _{veg and r2b_ratio} represent the index threshold _values . If any of R _{grn_veg} and R _{red_veg} is greater than or equal to the threshold value T _{veg, ratio} , it is removed as a vegetation region. Here _{, values of} T _{veg, DN} = 20, T _{veg, g2b_ratio} = 1.25, T _{veg, r2b_ratio} = 2.0, and T _{veg, ratio} = 0.3 were adopted.

(6th step)
Next, an index called a rectangular index for each region is calculated as follows.
(A) The first axis and the second axis are determined from the two-dimensional coordinates of a set of edges in the region (referred to as an edge group).
(B) Express the existence range of the area with the values of the first axis and the second axis, and multiply (maximum value-minimum value + 1) of the first axis and (maximum value-minimum value + 1) of the second axis. , Get the rectangular area surrounding the region.
(C) (Area area / Area area surrounding the area) is defined as a rectangle index.
Here, if the rectangular index falls below a certain value (for example, 0.4), it is determined that there is a high possibility that it is not a building, and is excluded from the extraction target.

A supplementary explanation of the rectangular index will be given. FIG. 5 shows a conceptual diagram of rectangular index calculation. The first and second axes are calculated from the edge group of a certain region, and the smallest rectangle as shown in the figure surrounding the target region is obtained from the rectangles whose sides are parallel to the first and second axes. In determining the first and second axes, the slope of the straight line is voted by using an edge group in which the distance between two points is a constant value (between 5 pixels and 20 pixels), and the slope at which the maximum frequency occurs is determined. The orientation was the first axis, and the second axis was determined to be orthogonal to the first axis.
i _dx = S _actual / S _rect (1)
It is defined as Here, i _dx represents the rectangular index, S _actual represents the area of a certain region, and S _rect represents the area of the rectangle surrounding the region. The rectangular index takes a value in the range of 0 to 1, and the closer to 1, the closer to the rectangle. With such an index, the degree of “closeness” in a shape close to a rectangle can be expressed simply and accurately as an index.

(7th step)
In step 7, using the rectangular index of equation (1), the roof / building divided by the shadow is integrated to estimate the original area.
For example, search for a region near a certain region A,
(A) The rectangular index when merged is improved over each rectangular index.
(B) The specified threshold value (0.7) or more.
(C) The difference in the average luminance value of the region is within a certain value (30).
It is determined whether or not the condition is satisfied. When satisfy | filling, let the area | region where the rectangle index | exponent at the time of merging becomes the largest among the adjacent area | regions as a candidate for merging, and let this be the area | region B temporarily. A similar search is performed in the region B. When the region A satisfies all the conditions (a) to (c) and the rectangular index at the time of merging is maximized, the regions A and B are merged with each other. .

  For example, as shown in FIG. 6, a rectangular index is calculated when all the regions β, γ, and δ adjacent to the target region α are integrated. In this case, the first and second axes are calculated using an edge group excluding the edge groups close to each other among the edge groups of the two areas. The axis is determined by voting the slope of a straight line using an edge group in which the distance between two points is a constant value (5 pixels or more and 20 pixels or less), and the slope with the highest frequency is taken as the direction of the first axis. The second axis was determined to be orthogonal to the first axis.

(8th step)
Next, regions of a certain area or more obtained under N _disc types of discretization widths are selected in descending order of the rectangular index. However, if any part of the area overlaps the already selected area, the area is not selected.

(9th step)
The area that was not selected is selected again. If the overlap with the already selected region is less than a certain value (30%) and less than a certain area (80 pixels), only the non-overlapping portion is additionally selected as a new region. However, only when the rectangular index is calculated for the non-overlapping portion and is equal to or greater than the threshold (0.45).

(10th step)
Finally, fill the holes in the area.

"Illustration"
Examples are shown below. FIG. 7 is a diagram showing a result of edge extraction in an urban area where low-rise buildings are densely gathered (hereinafter referred to as area 1). (A) is an aerial photograph in which an area having an actual size of about 75 m × 75 m is photographed with a photographing resolution of 300 × 300 pixels. (B) is a result of using a known Canny filter for the data of (a). (C) is an edge extraction result (first stage) when the discretization width 40 is used in the discretization described above. Further, (d) is an edge extraction result (second stage) when the discretization width 30 is used, and (e) is an edge extraction result (third stage) when the discretization width 20 is used.

  FIG. 8 is a diagram showing a result of edge extraction in an urban area where high-rise buildings and low-rise buildings are mixed (hereinafter referred to as region 2). (A)-(e) are the same as FIG. 7, (a) is an aerial photograph, (b) is a result of using a Canny filter, (c) is an edge extraction result when using the discretization width 40 (First stage), (d) is an edge extraction result (second stage) when the discretization width 30 is used, and (e) is an edge extraction result (third stage) when the discretization width 20 is used. is there.

  FIG. 9 is a diagram showing a result of edge extraction in an area where a tree is adjacent to a building (hereinafter referred to as area 3). (A)-(e) are the same as FIG. 7, (a) is an aerial photograph, (b) is a result of using a Canny filter, (c) is an edge extraction result when using the discretization width 40 (First stage), (d) is an edge extraction result (second stage) when the discretization width 30 is used, and (e) is an edge extraction result (third stage) when the discretization width 20 is used. is there.

  FIG. 10 is a diagram showing the result of edge extraction in an area where there are many buildings with a dormitory roof (hereinafter referred to as area 4). (A)-(e) are the same as FIG. 7, (a) is an aerial photograph, (b) is a result of using a Canny filter, (c) is an edge extraction result when using the discretization width 40 (First stage), (d) is an edge extraction result (second stage) when the discretization width 30 is used, and (e) is an edge extraction result (third stage) when the discretization width 20 is used. is there.

  FIGS. 11, 12, 13, and 14 show the results of region extraction in the above-described region 1 (FIG. 7), region 2 (FIG. 8), region 3 (FIG. 9), and region 4 (FIG. 10), respectively. FIG. In each figure of FIGS. 11-14, (a) is an aerial photograph of each area, (b) is a region extraction result when using a discretization width 40 (first stage), (c) is a discretization width 30 Region extraction result when used (second stage), (d) is an edge extraction result when using the discretization width 20 (third stage), and (e) is a result of three results (( It is a figure of the final result which integrated b)-(d)). In any of FIGS. 11-14, (e) becomes the state which took the best place of three results locally so that it may become clear when compared with each result of (b)-(d). Regions (edges) are extracted more accurately.

  FIG. 15 is a diagram of the region 1, where (a) to (f) are (a) aerial photographs, (b) reference points for comparing region extraction results, and (c) region extraction results (main) Principal component analysis is used when calculating the direction), (d) region extraction results (deciding by voting the slope of a straight line passing between two points when calculating the main direction), (e) ITT Visual Information Solutions These are the region extraction results when Scale = 50 and Merge = 50 are set in the image processing software ENVI EX, and (f) the region extraction results when Scale = 50 and Merge = 80 are set in ENVI EX. Note that in (c) to (f), the edge of the building is clearly displayed in the color image, but in this figure, the edge looks like the edging of the building.

  It should be noted that the parameters required to be set by the function called “Feature Extraction” of ENVI EX are “Scale Level” and “Merge Level”. In the target area of this example, there are many roofs with textures with large dispersion values. Changing the “Merge Level” changes the size and number of areas that will ultimately be left, so “Scale Level = 50.0” is the same, and two types of values, “Merge Level = 50.0” and “Merge Level = 80.0” Executed with.

  FIG. 16 is a diagram regarding the area 2. (a) to (f) are (a) aerial photographs, (b) reference points for comparing region extraction results, and (c) region extraction results (main) (Principal component analysis is used when calculating the direction), (d) Region extraction result (decided by voting the slope of a straight line passing between two points when calculating the main direction), (e) Scale = in ENVI EX (F) Region extraction result when Scale = 50, Merge = 80 is set in ENVI EX.

  FIG. 17 is a diagram for the region 3. (a) to (f) are (a) aerial photographs, (b) reference points for comparing region extraction results, and (c) region extraction results (mainly). (Principal component analysis is used when calculating the direction), (d) Region extraction result (decided by voting the slope of a straight line passing between two points when calculating the main direction), (e) Scale = in ENVI EX (F) Region extraction result when Scale = 50, Merge = 80 is set in ENVI EX.

  18A and 18B are diagrams for the area 4. (a) to (f) are (a) an aerial photograph, (b) a reference point for comparing region extraction results, and (c) region extraction results (mainly). (Principal component analysis is used when calculating the direction), (d) Region extraction result (decided by voting the slope of a straight line passing between two points when calculating the main direction), (e) Scale = in ENVI EX (F) Region extraction result when Scale = 50, Merge = 80 is set in ENVI EX.

  In each of FIGS. 15 to 18, (c) and (d) by region extraction of this embodiment, as is clear from comparison between the middle (c) and (d) and the lower (e) and (f). It can be seen that the rectangular area of the building is extracted more accurately. In ENVI EX (e) and (f), there are a large number of unnecessary edges and edges that are too fine. Compared with (c) and (d), the region extraction performance is poor.

<Discussion>
In the region extraction method according to the above embodiment, edges obtained by applying different discretization widths are integrated. This is essentially the same as conversion to different spatial resolutions and processing. However, the edge is preserve | saved unlike a simple smoothing filter by applying and integrating by changing an offset value. Most importantly, for regions obtained by applying different discretization widths, adopting a spatial smoothing parameter that is locally optimal to adopt sequentially from regions with a large rectangular index. It is equivalent to being.

  FIGS. 11 to 14 show that the optimum spatial scale parameter can be selected locally according to the size of the roof or building through the integration process. However, labeling using the discretized image takes time, and the processing time increases as the image size increases. For example, in the case of a computer using a 6 GB memory for a CPU operating at a clock of 3.2 GHz, the average processing time in each target area was about 90 seconds.

  Note that it is preferable to preferentially extract a rectangular-shaped region found on a roof or a building by using an index related to the shape calculated from the edge of the region. At that time, there are many roofs and buildings that do not completely close even if edges are extracted, which is a factor that lowers the accuracy of area extraction, but for edges that are not completely closed, the shape of a rectangle or triangle It is preferable to add a process of closing the edge when there is a high possibility.

  Further, by using, for example, three different discretization widths as described above, reference points 1-a, 1-b, 3- in FIGS. 15-18 are obtained from the result of Δd = 20 ((e) in FIGS. 7-10). It can be seen that the edges of the roof with shadows such as a and 4-a are clearly extracted. Even in this case, it can be confirmed that the image is not separated without the edge closing process, and the effectiveness of the edge complementation (concatenation) process can be confirmed. That is, it is considered that the area extraction accuracy has been improved by the effects of the two factors of edge complement processing and integration of results of different discretization widths. However, if the condition of “the difference in the average luminance value of the areas is within a certain value” shown in step 7 is not present, the result of excessively integrating the clearly separated roofs was confirmed.

  From the results of FIGS. 15 to 18, it can be seen that the triangular area seen on the dormitory roof is also well extracted. The ideal triangle rectangle index remains at 0.5 and is not preferentially adopted. A factor that has been successfully extracted this time is that the rectangular area around the triangular area has been successfully extracted. Depending on the discretization width and the offset value, the triangular area and the rectangular area may be integrated. However, when the region extraction results obtained with three different discretization widths are evaluated on a scale called the rectangle index, the rectangle index of the area where the triangle area and the rectangle area are integrated is lower than the rectangle index of the original rectangle alone area. It is hard to be adopted. Based on such a selection process, the rectangular area was first selected, and then the triangular area was extracted.

  Compared with the result of the existing region extraction method, the range of the region does not exist in the result of the above method. This is partly due to the upper limit on the area of the area, but it is largely due to the fact that if the rectangular index of the area is less than a certain value (0.4), the possibility of roofs / buildings is low. doing. Most but not all vegetation and the area including the cemetery in FIG. 15 are far from being rectangular in shape and can be removed as a result. In other words, it can be said that it functions sufficiently in terms of building area extraction.

The rectangular index will be described from the viewpoint of area. In this method, since the rectangle index is extracted sequentially, a relatively small area is often extracted preferentially, and an area that captures a rough outline is selected even if the rectangle index is slightly small. It may not be. In order to deal with applications that would like to capture such rough outlines, we examined the method. For example, by executing the correction using the following equation (2), a region corresponding to a plurality of roofs / buildings is obtained in some regions, which is larger than the results shown in FIGS.
Di _dx = C ₁ · ln (S _actual · C ₂ + C ₃ ) (2)

Here, Di _dx represents a correction value of the rectangular index, and C _{1 to} C ₃ represent coefficients determined empirically. As a result of empirical confirmation by applying to the above target area, C ₁ = 0.05, C ₂ = 100.0, and C ₃ = 1.0. Among the coefficients of equation (2), it was found that C ₁ has a great influence on the final extraction result. If C ₁ = 1.0 is set, the correction value is too large, and the rectangular index of the road and the area having a large area connected to it becomes high, so that it is preferentially selected accordingly. The number of buildings extracted decreased. This time, we are targeting the area where traditional Japanese buildings are lined up, but the functions and parameter values in this rectangular index correction should be adjusted with the building to be extracted in mind.

  On the other hand, in the case of calculating the rectangular index, FIGS. 15 to 18 (c) show the result of region extraction when the axis is determined by principal component analysis. At the reference points 2-a, 3-b, 3-c, and 4-b, when the principal component analysis was applied, the slate roof region was divided and partly extracted. If the principal component analysis is applied using the edge group unless it is a complete rectangle, the first principal component is not always parallel to the roof edge, especially in the roof that is divided by the shadow as the object of this time. No results were obtained. On the other hand, in the area extraction result using principal component analysis in 1-c, four roofs are extracted as one large area.

  Since it was confirmed that the region extraction result lacks stability by using principal component analysis in this way, the principal component analysis is not applied, and the slope of the straight line using an edge group in which the distance between two points is a constant value. The slope at which the maximum frequency occurs was defined as the direction of the first axis, and the second axis was determined to be orthogonal to the first axis. When the upper limit and lower limit of the distance between the two points were changed, the result of partial region extraction also changed, but the change was not as great as the result of principal component analysis. This threshold value should be determined empirically according to the characteristics of the target area.

  Finally, the removal of the vegetation area will be described. In the present embodiment, the purpose is to extract buildings in urban areas, and vegetation was a target to be removed. As a pre-process for region extraction, a method of removing pixels with high vegetation by referring to luminance values in pixel units is also conceivable. However, by removing the vegetation partially applied to the roof or road, the original shape of the roof or building may not be obtained, and it may be divided or the area may be too small to be extracted.

Therefore, as a result of the study, we decided to apply the process at the final stage, including the vegetation area as an area extraction target. It can be said that this policy has contributed to high-precision extraction in practice. For example, when trying to remove red vegetation, the red roof may be removed by mistake. Therefore, in order to carefully remove red vegetation, the threshold value was set high as T _{veg, r2b_ratio} = 2.0. With the same idea, shadows were extracted as one region, and when the rectangular index improved, it was dealt with so that it could be integrated with neighboring regions. The shadow-like area contains many roads and we considered removing it using luminance values. However, we confirmed that there were cases where roofs and buildings were removed by mistake, so we did not remove them.

  As described above, according to the region extraction of this embodiment, a building or a roof can be effectively extracted as a region in a densely built-up area. In this method, in order to effectively extract a roof having a texture with a large variance of luminance values, the luminance values are discretized into a small number of values, and regions having the same discretized values are labeled. Then, the rectangular index calculated from the edge of the area is used to preferentially extract the rectangular area seen on the roof or building. In addition, for the edge group that is not completely closed, a process of closing the edge when the possibility of a rectangular or triangular shape is high was added.

  In an experiment using a 25cm resolution aerial photograph in which an area where traditional buildings in Kyoto City are densely photographed, three different discretization widths were applied. Two factors are effective, and shadow regions that are difficult to extract with the existing method are clearly extracted. The most important feature of the method is that the region group obtained by applying a plurality of different discretization widths is adopted sequentially from the region with a large rectangular index, so that the spatially optimum This is the point that realizes the processing using the smoothing parameter.

  In addition, the triangular area seen on the dormitory roof has a low rectangular index and is not preferentially adopted, but as a result, it was successfully extracted. The rectangular index of the area where the triangular area and the rectangular area are integrated is less likely to be adopted than the rectangular index of the original single rectangle area, and the triangular area is extracted by selecting the rectangular area first. Finally, in all the target areas of urban areas where low-rise buildings are densely populated, urban areas where high-rise buildings and low-rise buildings are mixed, areas where high trees are adjacent to buildings, and areas where many buildings with dormitory roofs exist Better results were obtained than the region extraction method. Therefore, it was confirmed that the method characterized by the discretization of the luminance value and the utilization of the rectangular index is useful.

In the above embodiment, the aerial photograph is used as the original data, but it is also possible to use data of a photograph taken from an artificial satellite photographed with high accuracy instead of the aerial photograph.
In addition, it should be considered that the embodiment disclosed this time is illustrative and not restrictive in all respects. The scope of the present invention is defined by the terms of the claims, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

10: Computer device (region extraction device)

Claims (7)

An area extraction method for extracting an area of a building based on photo data taken from an aircraft or an artificial satellite,
A plurality of different discretization widths are set, and for each discretization width, the luminance value of the data is discretized into a plurality of values discretely set by the discretization width,
Connect the pixels with the same value in the discretized image obtained by discretization, extract a region of a shape close to a rectangle as a candidate for the region of the building,
Of the plurality of region groups extracted for each of the plurality of different discretization widths, a region having a shape closer to a rectangle is adopted as a region of the building.   The region extraction method according to claim 1, wherein a rectangle index defined by (area of area / area of rectangle surrounding area) is used as an index representing a shape close to a rectangle.   3. The region extraction method according to claim 1, wherein in the extraction, when adjacent regions are merged with each other, the merger can be executed when the regions become more rectangular.   The area extracting method according to claim 2, wherein when the rectangular index is smaller than a predetermined value, the area is not adopted as a building area.   The region extraction method according to claim 1, wherein a region estimated to be vegetation is excluded from extraction targets based on the RGB luminance values. An area extraction program for extracting an area of a building based on data of a photograph taken from an aircraft or an artificial satellite,
A function of setting a plurality of different discretization widths and discretizing the luminance value of the data into a plurality of values discretely set in the discretization width for each of the discretization widths;
A function of connecting pixels having the same value in a discretized image obtained by discretization and extracting a region having a shape close to a rectangle as a candidate for a region of the building; and
A region extraction program for causing a computer to realize a function of adopting a region having a shape closer to a rectangle as a region of a building among a plurality of region groups extracted for each of the plurality of different discretization widths. An area extraction device that extracts an area of a building based on photo data taken from an aircraft or an artificial satellite,
A function of setting a plurality of different discretization widths, and for each discretization width, discretizing the luminance value of the data into a plurality of values discretely set in the discretization width;
A function of connecting pixels having the same value in a discretized image obtained by discretization and extracting a region having a shape close to a rectangle as a candidate for a region of the building;
A region extracting apparatus having a function of adopting a region closer to a rectangle as a region of a building among a plurality of region groups extracted for each of the plurality of different discretization widths.