JP2013089021A - Image processing unit and image processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To automatically determine a threshold to extract regions with different wavelength ranges.SOLUTION: An image processing unit includes: a normalizing index calculation part for calculating a normalizing index to identify a pixel in pixels in geographical image data taken from a high altitude, on the basis of a spectral luminance value of a first wavelength range and a spectral luminance value of a second wavelength range different from the first wavelength range; a threshold determination part for obtaining a relationship between the normalizing index and an appearance frequency in each pixel, and obtaining the normalizing index that maximizes a degree of separation showing a proportion of in-class dispersion to inter-class dispersion using a discriminant analysis method to determine the obtained normalizing index as a threshold for identifying the pixel; and a region extraction part for extracting a region of image data to be distinguished by the first and second wavelength ranges, on the basis of the normalizing index corresponding to the threshold.

Description

  The present invention relates to an image processing technique, and more particularly to a technique for automatically separating a water area and a land area from an image such as a satellite image taken from a high sky.

  Research has been conducted on techniques for analyzing shallow water depths from the satellite image so that the bottom of the water is visible. In recent years, satellite images have improved performance in position accuracy, spatial resolution, spectral resolution, and the like, and the accuracy of water depth obtained as an analysis result has also improved. When applying the method for estimating the water depth, it is necessary to extract only the water area from the satellite image. In order to extract the water area, information on the boundary line (coast line) that separates the water area from the land area is necessary. If information that separates the water area from the land area is obtained, the land area can be extracted at the same time.

  Conventionally, extraction of water areas can be performed by manually entering boundary lines to separate land areas from water areas, or by setting threshold values for index images that have been subjected to image processing that emphasizes water areas or land areas. I went. In Patent Document 1 below, in order to binarize an input image with a predetermined fixed threshold, a band in the RGB image is separated into a water area and a land area by the fixed threshold.

Japanese Patent Publication No. 2005-148906

  When a method for manually determining a fixed threshold value for water area extraction is used, there is a problem in that the result depends on the operator because the determination criteria differ depending on the operator. In addition, since the threshold value is manually determined, there is a problem that the processing time increases in proportion to the area of the target area and the complexity of the coastline.

  An object of the present invention is to automatically determine a threshold value for extracting regions having different wavelength bands.

  In order to achieve the above object, the present invention creates a moisture index image and a histogram from a satellite image, and automatically determines a threshold value for separating a water area and a land area from the histogram by a discriminant analysis method.

  According to one aspect of the present invention, a spectral luminance value in a first wavelength band and a spectral luminance value in a second wavelength band different from the first wavelength band in a pixel in geographic image data taken from a high sky, , A normalization (water) index calculation unit for calculating a normalization (water) index value for identifying a pixel based on, and a relationship between the normalization (water) index value and the appearance frequency in each pixel, A threshold determination unit that obtains a normalized (water) index value that maximizes the degree of separation, which is the ratio of intra-class variance and inter-class variance, by discriminant analysis, and determines the threshold value for binarization for identifying the pixel And a region extracting unit that extracts a region of image data distinguished by the first and second wavelength bands based on a normalized (water) index value corresponding to the threshold value. An image processing apparatus is provided.

  The above apparatus can automatically determine a threshold value for extracting regions having different wavelength bands.

  Further, a micro area removing unit that uses a micro area having a different wavelength existing in the area of the image data distinguished by the first and second wavelength bands as a pixel having a spectral luminance value in the area of the image data. It is preferable to have.

  The threshold value determination unit obtains a relationship between the normalized (water) index value and the appearance frequency in each pixel as a normalized (water) index value histogram, and includes a histogram of the intra-class variance and the inter-class variance. When the normalized (water) index value that is the scanning start point at which the frequency is maximum is negative, the scanning direction is the positive direction, and when the normalized (water) index value that is the scanning start point is positive, scanning is performed. The direction is preferably a negative direction. As a result, the starting point is closer to zero than -1 and +1, and the scanning range is narrower than that of a general discriminant analysis method starting from -1 or +1, which is efficient.

  The threshold value determination unit creates the normalized (water) index value histogram excluding the minute area to be ignored extracted immediately before, and applies a discriminant analysis method to the normalized (water) index value histogram to set the threshold value. The threshold value may be determined by repeating the determining process. Thereby, in the process using the satellite image, it is possible to automatically determine a more appropriate threshold value of the moisture index value in consideration of factors that affect the result of the discriminant analysis method.

  According to another aspect of the present invention, a spectral luminance value in a first wavelength band and a spectral luminance value in a second wavelength band different from the first wavelength band in a pixel in geographic image data taken from a high sky. And a normalization (water) index calculation step for calculating a normalization (water) index value for identifying a pixel based on the relationship between the normalization (water) index value and the appearance frequency in each pixel. , A threshold value for determining a normalized (water) index value that maximizes the degree of separation, which is the ratio between the intra-class variance and the inter-class variance, by a discriminant analysis method, and determining it as a threshold value for binarization for identifying the pixel A determination step; and a region extraction step of extracting a region of image data distinguished by the first and second wavelength bands based on a normalized (water) index value corresponding to the threshold value. An image processing method is provided.

  The present invention may be a program for causing a computer to execute the above-described image processing method, or a computer-readable recording medium for recording the program.

  According to the present invention, it is possible to automatically and accurately determine a threshold value for performing binarization for demarcating boundaries between regions having different wavelength bands, so that work efficiency is improved.

It is a functional block diagram which shows the example of 1 structure of the image analysis apparatus by one embodiment of this invention. It is a figure which shows an example of the satellite image containing a water area and a land area. FIG. 3 is a histogram of NDWI for the image of FIG. 2. It is a flowchart figure which shows the flow of the water area extraction process which does not consider a micro water area among the water area extraction processes by the 1st Example of this invention. It is a figure which shows an example of the satellite image after a water area extraction process. In the water area extraction processing according to the second embodiment of the present invention, the NDWI histogram including the minute water area to be ignored and the NDWI histogram not including the minute water area to be ignored are compared with the satellite image of FIG. FIG. It is a flowchart figure which shows the flow of the water area extraction process which considered the micro water area among the water area extraction processes by the 2nd Example of this invention.

  Hereinafter, an image processing apparatus according to an embodiment of the present invention will be described by taking a satellite image (geographic image) process including a water area and a land area as an example.

  FIG. 1 is a functional block diagram showing a configuration example of an image processing apparatus according to the first embodiment of the present invention. The image processing apparatus illustrated in FIG. 1 includes a processing device 10, a storage device 20, and an input / output device 30. The processing device (CPU) 10 includes an image processing unit 40. The image processing unit 40 includes an NDWI (Normalized Difference Water Index) calculation unit 100, a threshold value determination unit 200, and a water area extraction unit. 300 and a minute region removing unit 400. The storage device 20 receives satellite image data, and includes a satellite image data storage unit 21, an NDWI arithmetic expression storage unit 22, a threshold storage unit 23, and a water area data storage unit 24. Further, it is general to store a program for causing the CPU 10 to execute processing. The input / output device 30 includes an input device 31, a display device 32, and a printer 33. In addition, it is common to have a communication interface.

  FIG. 2 is a satellite image 501 to be processed, in which a water area 502 and a land area 503 and a coastline 504 that is a boundary between the water area and the land area are displayed. A water area 502 is extracted from the image 501.

  FIG. 3 is a diagram illustrating an example of a moisture index histogram obtained by converting a target image. The moisture index handled in this specification is an index generally called NDWI, for example. NDWI is a value obtained based on the following equation (1) from a value obtained by converting a value of a plurality of wavelength bands given in each pixel such as a satellite image into a spectral radiance value having an energy unit. is there.

    NDWI = (Spectral radiance value in short wavelength band-Spectral radiance value in long wavelength band) / (Spectral radiance value in short wavelength band + Spectral radiance value in long wavelength band) (1)

  Here, the wavelength is the wavelength of sunlight reflected by the ground surface, which is observed by an optical sensor. The human eye can feel the wavelength band of visible light, and recognizes light with a short wavelength as blue light within the wavelength range of visible light, and recognizes light with a long wavelength as red light within the wavelength range of visible light. To do. Artificial satellites equipped with optical sensors (referred to as optical satellites) have different observable wavelengths depending on the optical sensor, but sensors mounted on many optical satellites are close to the three primary colors of light, Blue (blue), Green ( It is a sensor that observes four bands of NIR, which is the near-infrared wavelength band and green (Red). The calculation formula for the spectral radiance value varies depending on the sensor to be photographed.

  For specific NDWI calculations, for example, an image taken by the WorldView-2 satellite equipped with an 8-band sensor operated by DigitalGlboe in the United States is the spectral radiance value and length of the Coastal band in the short wavelength band. Substitute the spectral radiance value of the NIR2 band for the wavelength band. Similarly, images taken by satellites equipped with 4-band sensors such as QuickBird operated by DigitalGlobe in the US and GeoEye-1 and IKONOS operated by GeoEye in the US are spectral radiances in the blue band in the short wavelength band. Substitute the spectral radiance value of the NIR band for the value and long wavelength band.

  NDWI is said to have a greater surface water content as its value increases, and is used to set a threshold and separate the land area from the water area. In addition, NDWI takes values from −1 to +1 from Equation (1).

  In the case of an image showing a land area and a water area, NDWI is calculated from, for example, all pixels of the image, and the histogram obtained by counting the appearance frequency (corresponding to the number of pixels) of NDWI largely forms two mountains (601). Pixels in the land area are concentrated on the mountain (602) in the negative number range of NDWI, and pixels in the water area are concentrated on the other mountain (603) in the higher positive number range. At this time, the optimum threshold value of NDWI for separating the water area and the land area is a valley between two mountains (broken line indicated by 604).

  FIG. 4 is a flowchart showing the flow of the water area extraction process according to the present embodiment. First, the NDWI calculation unit 100 of the image processing unit 40 converts image data (target image) input by the input / output device 30 and stored in the satellite image data storage unit 21 into a spectral radiance value, for example. Then, an NDWI image (NDWI value for each pixel) is calculated by the calculation formula (1) stored in the NDWI calculation formula storage unit 22 (step S1). The image value handled here is a signed floating point number. Next, the NDWI value in the image showing both the water area and the land area as shown in FIG. 2 has a characteristic that the appearance frequency is largely biased into a positive area and a negative area as shown in FIG. Form a mountain. At this time, the optimum NDWI value for separating the two regions is between two mountain valleys. This valley, that is, the threshold value of NDWI is determined by a discriminant analysis method.

The threshold value determination unit 200 determines a threshold value by a discriminant analysis method (step S2).
Discriminant analysis is also called binarization of Otsu. This is a method for obtaining a threshold value that maximizes the degree of separation, which is the ratio of intra-class variance and inter-class variance, assuming a histogram that forms two normal distribution peaks (classes). The threshold at this time is a trough between the two histogram peaks.

  A general discriminant analysis method is a method of scanning from the minimum value to the maximum value or the reverse direction of the index. Here, the starting point and the direction of efficient scanning are determined by utilizing the fact that the range of the histogram is from −1 to +1 and the fact that there is one peak in almost the positive and negative ranges. More specifically, the NDWI that is the starting point of scanning is first set to NDWI having the highest frequency (603 in FIG. 3), and is specified when the histogram is generated. This origin is the NDWI that points to the highest peak of the two peaks in the histogram. Next, when the starting NDWI is negative, the scanning direction is the positive direction. On the other hand, when the starting NDWI is positive, the scanning direction is the negative direction. As a result, the starting point is closer to zero than -1 and +1, and the scanning range is narrower than that of a general discriminant analysis method starting from -1 or +1, which is efficient.

  Note that the NDWI histogram does not form a clear valley as shown in FIG. 3 when an image having a large water area relative to the land area or an image having a large land area relative to the water area is similarly targeted. There is. The discriminant analysis method may not be able to determine an expected threshold value for a histogram without a clear valley. However, by using the characteristics of a histogram having a target threshold value near zero of NDWI, it is possible to stably determine the NDWI threshold value by looking closely at the vicinity of zero of NDWI even when the valley is difficult to understand. it can. The determined threshold value is stored in the threshold value storage unit 23.

  All pixels in the original image and the NDWI image correspond uniquely. The water area extraction unit 300 compares the corresponding NDWI with the NDWI threshold value stored in the threshold value storage unit 23, regards a pixel having an NDWI lower than the NDWI threshold value as a land area, and assigns all the bands of the pixel to values that are ignored. By doing so, the pixel value of the land area is deleted. As a result, only water area information is extracted from the image. As an example, an inevitable value of −999 is assigned to all the bands to be ignored (step S3).

  The water area extracted image 501 based on the NDWI threshold is separated into a water area 502 and a land area 503, and a coastline 504 appears. The water area is stored in the water area data storage unit 24.

  However, in actual terrain, a small water area (micro area) that is not a target shallow water area, such as a paddy field or a pond, is detected in the extracted water area image. Therefore, the minute area removing unit 400 considers the land area by substituting a value that can be ignored as the land area into the minute water areas 505a and 505b. At this time, first, an 8-connected labeling processing algorithm for 3 × 3 pixels is applied to the image extracted from the water area. In general, the 8-connected labeling process is applied to a binarized image. For example, when 1 pixel is labeled in an image binarized with 0 and 1, 1 pixel existing in the vicinity of 8 around the 1 pixel is set as the same group. As a result of processing for all one pixels, a plurality of groups are formed, and one pixel always belongs to one group. Here, the processing is performed by regarding the image as a binarized image of two types of pixels, that is, a pixel in which a value to be ignored for all the bands is substituted, and other pixels, thereby performing an 8-connected labeling process. The number of pixels belonging to is counted. For example, a group with the number of belonging pixels less than half of the group with the largest number of belonging pixels is regarded as a small water area to be ignored, and -998 is substituted for all the bands of pixels belonging to the group (step S4). Note that step S4 is an arbitrary process, and the process may be terminated up to step S3.

  Through the above processing, the water area of the satellite image can be extracted. The water area can be output by the display device 32, the printer 33, or the like.

Next, a satellite image processing technique according to the second embodiment of the present invention will be described.
As described in the first embodiment, the expected water extraction can be realized by performing the processing shown in FIG. 4, but the NDWI histogram 601 (FIG. 3) used when determining the threshold value is used. Since the above-mentioned minute water areas that should be ignored are included, if there are so many minute water areas that should be ignored that the NDWI histogram is affected, the NDWI threshold is set so that the expected high-precision water area extraction cannot be performed. May affect the decision process. In the present embodiment, processing in consideration of this point is performed. 6 is a diagram corresponding to FIG. 3, and FIG. 7 is a diagram corresponding to FIG.

  First, the NDWI image generation unit 100 converts the input image data (target image) into a spectral radiance value, and further generates an NDWI image (step S11). Then, a 1-band grayscale image having the same size as the target image is created as a mask image. At this time, an initial value is assigned to each pixel. Here, −1 is set as a specific example (step S12). Further, the work image is generated by copying the target image (step S13).

  The target image, the NDWI image, the mask image, and the work image have the same size, and each pixel in each image uniquely corresponds. Here, the threshold value determination unit 200 creates an NDWI histogram using only pixels in the NDWI image corresponding to pixels that are not 998, which is a value representing a minute water area to be ignored in the mask image. Immediately after generating the mask image, all the bands of all the pixels are −1, so that the NDWI histogram 605a is created using all the pixels for the first time. When the land area and the water area are reflected, the pixels of the land area are concentrated on the mountain 602a in the negative range of the NDWI, and the pixels of the water area are concentrated on the mountain 603a having the higher positive number range. The threshold of the NDWI that separates the optimal water area from the land area is a valley 604a between two mountains. The NDWI threshold value is determined by the discriminant analysis method as in FIG. 4 (step S2) (step S14).

  All pixels in the working image and the NDWI image uniquely correspond to each other. Compare the corresponding NDWI with the NDWI threshold, consider pixels with NDWI lower than the NDWI threshold as the land area, substitute all the bands of the corresponding pixels in the working image pixels into values that are ignored, and erase the pixel values in the land area To do. As a result, only water area information is extracted from the image. Note that, for example, -999 is substituted for all the bands to be ignored (step S15).

  The water area extracted image 501 based on the NDWI threshold is separated into a water area 502 and a land area 503, and a coastline 504 appears. However, in the extracted water area image, small water areas 505a and 505b that are not targeted shallow water areas, such as paddy fields and reservoirs, are detected. Therefore, as in FIG. 4 (step S4), a value that ignores a minute water area as a land area is substituted. Here, as a specific example, -998 is substituted for all bands of pixels corresponding to a minute water area that should be ignored for the work image (step S16).

  Up to this point, the work image is represented by −999 representing the pixel value of the water area and the pixel value of the land area, and −998 representing the minute water area to be ignored. The corresponding pixels of the working image and the mask image are compared, and the number of pixels in which one pixel is a water area and the other pixel is outside the water area is totaled. When the number of pixels that meet this condition is zero, the water area on the work image matches the water area on the mask image. If the water area on the work image matches the water area on the mask image (Yes in step S17), the process ends (end). The work image at this time is a result image obtained by extracting the water area. If the number of pixels meeting the conditions is not zero (No in step S17), the work image is set as a mask image (step S18). Then, the target image is copied again to generate a work image (return to step S13), and the same processing is repeated.

  Here, for example, immediately after the mask image generation (step S12), since all the bands of all the pixels on the mask image are −1, the number of pixels that meet the condition does not become zero. Therefore, the work image is set as the mask image, and the work image generation (step S13) is performed again. The threshold determination by the second and subsequent discriminant analysis methods (step S14) forms an NDWI histogram (601a) excluding the minute water area that should be ignored just before extraction (from 605a to 601a). A discriminant analysis method is applied to the NDWI histogram (601a) excluding the minute water area to be ignored, and the NDWI threshold value is determined in consideration of the influence of the minute water area to be ignored. Here, the peak of the water area approaches from 605a to 601a, and the threshold also approaches from 604a to 606a.

  In the subsequent processing, similarly to the previous time, extraction of the water area by applying the threshold and update of the work image (step S15), the minute water area is excluded from the water area, and the work image is updated (step S16). If the water area pixel on the work image matches the water area pixel on the mask image, the process ends. If not, the process is repeated again using the work image as the mask image (steps S17 and S18). . Eventually, the influence of the minute area is reduced, and the threshold value is also changed from 604a to 606a.

  In step S17, the water area pixels on the working image and the water area pixels on the mask image may not be completely matched, and the condition may be relaxed if they are equal to or less than a few percent of the number of water area pixels on the working image.

  By using the above method, the boundary line between the water area and the land area, or the threshold value for determining the boundary line can be uniquely determined without depending on the judgment criteria of the operator. In addition, since most of the processing can be performed by a computer, work efficiency can be greatly improved.

  Furthermore, in the present invention, by adopting a moisture index having a finite range of ± 1, histogram processing on a computer can be simplified and made efficient.

  In the above-described embodiment, the configuration and the like illustrated in the accompanying drawings are not limited to these, and can be changed as appropriate within the scope of the effects of the present invention. In addition, various modifications can be made without departing from the scope of the object of the present invention.

  For example, instead of NDWI (Normalized Difference Water Index), normalized soil index (NDSI) and normalized vegetation index (NDVI) are used to extract land areas and forest areas. be able to.

  The present invention can be used in an image processing apparatus.

DESCRIPTION OF SYMBOLS 10 ... Processing apparatus, 20 ... Memory | storage device, 30 ... Input-output device, 40 ... Image processing part, 100 ... NDWI calculating part, 200 ... Threshold determination part, 300 ... Water area extraction part, 400 ... Minute area removal part, 21 ... Satellite Image data storage unit, 22 ... NDWI arithmetic expression storage unit, 23 ... Threshold storage unit, 24 ... Water area data storage unit, 31 ... Input device, 32 ... Display device, 33 ... Printer, 501 ... Target satellite image, 502 ... Water area, 503 ... land area, 504 ... coastline, 505a and 505b ... minute water area, 601 ... NDWI histogram, 602 ... range based on land area, 603 ... range based on water area, 604 ... histogram based on threshold Valley, 601a: NDWI histogram excluding minute water area, 602a: Range mainly composed of land area, 603a ... Range mainly composed of water area, 604a ... Minute water Threshold histogram valley which, 605a ... NDWI histogram containing micro waters, valleys of the histogram to the threshold when excluding 606a ... small waters when including.

Claims (6)

To identify a pixel in a geographic image data photographed from a high sky based on a spectral luminance value in a first wavelength band and a spectral luminance value in a second wavelength band different from the first wavelength band A normalization index calculation unit for calculating the normalization index value of
The relationship between the normalized index value and the appearance frequency in each pixel is obtained, and the normalized index value that maximizes the degree of separation, which is the ratio between the intra-class variance and the inter-class variance, is obtained by discriminant analysis, and the pixel is identified. A threshold value determination unit that determines a threshold value for binarization
A region extracting unit that extracts a region of image data distinguished by the first and second wavelength bands based on a normalized index value corresponding to the threshold;
An image processing apparatus comprising:   Further, a micro area removing unit that uses a micro area having a different wavelength existing in the area of the image data distinguished by the first and second wavelength bands as a pixel having a spectral luminance value in the area of the image data. The image processing apparatus according to claim 1, further comprising:   The threshold value determination unit obtains a relationship between the normalized index value and the appearance frequency in each pixel as a normalized index value histogram, and the frequency is maximized in the histogram of the intra-class variance and the inter-class variance. The scanning direction is a positive direction when the normalization index value serving as a scanning start point is negative, and the scanning direction is a negative direction when the normalization index value serving as a scanning start point is positive. Item 3. The image processing apparatus according to Item 1 or 2. The threshold determination unit
Creating the normalized index value histogram excluding the minute area to be ignored extracted immediately before, applying a discriminant analysis method to the normalized index value histogram histogram, and repeating the process of determining the threshold value to determine the threshold value The image processing apparatus according to claim 3. To identify a pixel in a geographic image data photographed from a high sky based on a spectral luminance value in a first wavelength band and a spectral luminance value in a second wavelength band different from the first wavelength band A normalization index calculation step for calculating the normalization index value of
The relationship between the normalized index value and the appearance frequency in each pixel is obtained, and the normalized index value that maximizes the degree of separation, which is the ratio between the intra-class variance and the inter-class variance, is obtained by discriminant analysis, and the pixel is identified. A threshold value determining step for determining as a threshold value for binarization,
A region extracting step of extracting a region of image data distinguished by the first and second wavelength bands based on a normalized index value corresponding to the threshold;
An image processing method comprising:   A program for causing a computer to execute the image processing method according to claim 5.

Images