JPS6320579A - Extraction processing system for feature amount of image - Google Patents

Abstract

PURPOSE:To improve the extraction accuracy of a fine object in an image without increasing the processing time, by comparing the object with a reference shape feature amount in a reference shape feature amount file, only with regard to a candidate object. CONSTITUTION:A satellite image 21 is inputted to an image analysis processing part 22, the processing is divided into two stages and executed, and in the first stage, a rough selection processing for extracting an object candidate point is executed while extending over the whole image. This processing is a processing extending over the whole image, therefore, it does not take a calculation time. For instance, the rough shape extracting parameter of an area, etc., is used. It is derived in advance with regard to a known object, and the rough shape extracting parameter is stored in advance in a reference shape parameter file 23 for rough selection. Subsequently, only with regard to each candidate point after a rough selection is processed, a fine specific shape object is extracted by a precise shape extraction processing using a Fourier descriptor. After the precise shape extraction processing is executed, the extracted fine shape object is outputted to an image display device 25.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an image feature extraction processing method, and more particularly to an image feature extraction processing method suitable for extracting a specific shape from an image. be.

[Conventional technology]

Conventionally, as a processing method for extracting shape features from images, Hideyuki Tamura, "Introduction to Computer Image Processing" (Souken Publishing, 1993)
March 1985)”, pages 87-94,
A process of calculating shape feature amounts for a binarized image, comparing each feature amount, and extracting a specific shape has been performed.

[Problem that the invention seeks to solve]

In the above-mentioned processing, in order to extract a minute object, it is necessary to calculate detailed feature amounts for each candidate shape, which has the problem of increasing processing time.

The present invention has been made in view of the above circumstances, and its purpose is to solve the above-mentioned problems in conventional image processing devices, and to extract fine objects in images without increasing processing time. An object of the present invention is to provide an image processing method that can improve accuracy.

[Means for solving problems]

The above objects of the present invention include: an image binarization processing unit; a shape feature calculation processing unit that calculates a shape feature from the binarized image;
An image analysis system having a reference shape feature amount file, and a shape time mu comparison detection processing section that compares the shape feature amount calculated by the shape feature amount calculation processing section and the reference shape feature amount in the reference shape feature amount file. In the above, the shape feature amount calculation processing section is constituted by a coarse shape feature amount calculation processing section and a precise shape feature amount calculation processing section, and the shape feature amount comparison detection processing section is constituted by a coarse shape feature amount comparison detection processing section. and a precise shape feature comparison detection processing section, and after selecting a candidate object by the coarse shape feature calculation processing section from the image binarized by the binarization processing section, This is achieved by an image feature extraction processing method characterized in that the precise shape feature comparison and detection processing unit performs a comparison with a reference shape feature in the reference shape feature file.

[Effect]

In this invention, image analysis! This is a two-step method in which candidate objects are selected using (macroscopic) shape features, such as circularity, circumference, and moments, and then precise shape features are calculated from only the candidate objects.

The above macroscopic shape feature amount processing is performed as follows.

(1) Binarize the projected image and label closed surfaces, that is, connected regions.

(2) Calculate macroscopic shape features for the labeled object. Here, as the macroscopic shape feature amount, use is made of area, 9 perimeter, 2 circularity, moment, etc., which require a short processing time.

(3) Perform threshold processing on macroscopic shape features,
Objects that are clearly different from those to be extracted are not included as candidates for precise shape feature processing.

As described above, the number of input candidate points for precise shape feature processing, which requires processing time, can be reduced by performing threshold processing on the entire captured image using feature quantities that require a short calculation time.

The next precise shape feature amount processing is performed in the following flow.

(1) Perform an enlargement process on the candidate object in consideration of the sensor aperture characteristics.

(2) Perform binarization and labeling.

(3) Calculate precise shape features for the labeled object. As the precise shape feature amount, a Fourier descriptor having fine structure information of the object is used.

(4) Perform threshold processing on the precise shape features to extract the target object.

As described above, by reducing the number of candidate points in advance and performing threshold processing using precise shape features, it is possible to improve extraction accuracy without increasing processing time.

〔Example〕

Embodiments of the present invention will be described in detail below with reference to the drawings.

The embodiment described below is an image processing system that extracts an object having a specific shape, for example, an H-shaped building, from a satellite image.

The building in the satellite image above is at most several pixels square (TM
In order to identify the shape as an H-shaped building, it is necessary to extract the object with an accuracy of about 1 pixel.

FIG. 2 shows an example of the configuration of an image analysis system for extracting a specific shape from a minute image, which is an embodiment of the present invention. In the figure, 21 is a satellite image file, 22 is an image analysis processing unit, 23 is a reference shape parameter file for rough selection, 24 is a reference shape parameter file for fine selection, 2
5 indicates an image display device.

The outline of the operation of this system is as follows. First, the satellite image 21 is input to the image analysis processing section 22. The processing is performed in two stages. The first stage is a process of extracting target candidate points over the entire image, and this is called a set selection process. Since this processing applies to the entire image,
A parameter for rough shape extraction, such as area, which does not require calculation time, is used. The rough shape extraction parameters are determined in advance for a known object and stored in the rough selection reference shape parameter file 23.

Next, only for each candidate point after the set selection process, a fine specific shape object is extracted by a precise shape extraction process using a Fourier descriptor. The reference shape parameters for precision selection are stored in the precision shape parameter file 24.

After the precise shape extraction process, the extracted fine shape object is output to the image display device 25.

The functions of the image analysis processing section 22 will be described in detail below with reference to FIG.

(1) Group selection based on macroscopic shape features After binarizing the input satellite image 11 (processing 12) and labeling the connected regions, the macroscopic shape parameters (features) of each region are
(Process 13), where the macroscopic shape parameters are: Area: A Circularity: P"/A (P: Perimeter) Moment: MiJ=硲Σ(X-x)'(y- y)'J (x y V): Coordinates of the center of gravity of the object (pixel,
line) lyJ: order of moment in each pixel direction, line direction (x, y): total coordinates within the object (pixel, line), etc.

Thresholds are set for each of these feature quantities and target candidates are detected (process 14). For, area: more than a pixel, a
Circularity of 1 pixel or less: b0 or more, b1 or less quadratic moment 20□ or more, 06 or less, etc.
It is stored in the rough selection reference shape parameter file 23.

The above-mentioned processing is performed on the binarized image that has been completely labeled, and only the shapes that satisfy the vA value are outputted from the object candidate detection 14.

In order to perform precise selection of the output result image based on the 0th-order enlarged image shown in FIG. 3(a), only the same labels are surrounded by a frame. This is the enlargement area setting process 15. Figure 3 (
In a), it is shown that five candidates were extracted by group selection.

(2) Precise selection based on enlarged image After narrowing down the target candidate area by the above processing, input image 11
Of these, the portion within each frame area is expanded by cubic convolution or the like (processing 16).

Binarization is performed on the enlarged image (processing 17). At this time, as shown in FIG. 3(b), the following process is performed in order to leave only the target object A inside the frame IA which is an enlarged version of the frame 1.

(2) Removing the object C that is in contact with the frame IA 9 For example, it is determined whether the (x, y) coordinates of the points forming each object are on each side of the frame IA.

(2) Next, in order to remove object B, a frame is set that circumscribes all the objects in frame IA, and only the object with the largest area is selected within the frame.

Eventually, object A is selected (FIG. 3(C)).

A Fourier descriptor for the selected object is calculated (shape parameter calculation process 18).

Fourier descriptor μ. tanyt)n is calculated like a pine.

Here, m: Number of vertices of a polygon formed by a closed curve representing the object π: Pi ratio n: High frequency component designation number of Fourier coefficient (n = 1.2...) L: Number of vertices of a polygon formed by a closed curve representing the object Perimeter length ΔΦ +flk: See Figure 4 Next, the reference shape parameter file 2 for f# dense selection
4. Compare and judge the training data stored in advance with the Fourier descriptor calculated above (process 1g)
.

For example, 5 calculated Fourier descriptors an, bn (n=
1.2. ...5), the Fourier descriptor of the training data is aTn* btn(n''1+2+...
.

・Calculate the Marlanobis distance D2 with 5).

D"=(X-y)Σ-tx-y)...(4)
Here, It is determined that The threshold value D"?H is set to an optimal value through experiments (
Object detection processing 19).

Images determined to be in the same class based on the threshold value are displayed as output images 20.

According to this embodiment, shape classification accuracy can be improved by calculating the fine shape feature amount of the object in addition to macroscopic shape analysis. Further, by performing macroscopic shape analysis as the first step, the target area for the second step processing, which requires a large amount of calculation, is reduced, which is effective in shortening the overall processing time.

Next, a method for efficiently creating training data in the reference shape parameter file 21 and a method for using the same, which were also used in the above embodiments, will be explained by showing an example.

The example described below focuses on the property that the Fourier descriptor used as training data can express the details of the shape as the number of terms increases, and the Fourier descriptor used as the training data is modified according to the resolution of the image to be analyzed. This is so that up to the number of terms are extracted and used as training data.

FIG. 5 shows a second embodiment of the present invention, which is an image analysis processing system that automatically extracts specific ground objects, such as buildings, from satellite images taken by different sensors.
In the figure, symbols 21 and 25 indicate the same components as previously shown in FIG. 2, 22A indicates an image analysis processing section, and 26 indicates a model graphic Fourier descriptor file.

The outline of the operation of this system is as follows. First, the satellite image 21 is read into the image analysis processing section 22, and the user inputs the resolution information of the input image from the image display device 25. The Fourier descriptor coefficients (training data) of the model figure that match the input resolution information are selected from the file 26. Next, by comparing the Fourier descriptor calculated from the input image with the training data. , automatically extracts shapes that match the training data from the input image.

Here, shape extraction from input images with different resolutions can be handled by selecting the number of terms of Fourier descriptor coefficients to be selected from the Fourier descriptor file 26 of the model figure.

Hereinafter, the functions of the above-mentioned image analysis processing section 22A will be explained in detail with reference to FIG. 6.

The input satellite image is binarized (processing 12), and the Fourier descriptor μ is applied to the binarized image. Calculate jantbn (
Process 18) 0 Next, in the feature space distance comparison process 19, the Marulanobis distance D2 is calculated and used as the feature space distance.

The above comparison process requires two steps.

(1) Model figure Fourier descriptor calculation process 27 Calculates the Fourier descriptors of a plurality of model figures 28 and stores the results in the model figure Fourier descriptor file 260 For example, the model figure shown in FIG. 7(a) The Fourier descriptor table of is made into a table up to N=100, as shown in FIG. 8(a).

Note that FIGS. 7(a) to (c) respectively show a model figure, an image obtained by a high-resolution sensor, and an image obtained by a low-resolution sensor.

(2) Fourier descriptor selection process 29 A Fourier descriptor is selected according to the resolution of the sensor. Here, up to the 10th term of the Fourier descriptor having the reproduction ability corresponding to the model figure shown in FIG. 7(b) is selected as the image taken by the high-resolution French HRV sensor.

After the above processing, the vector , it is determined that it belongs to the model figure class.

Also, a sensor image different from the above, for example, LA in the United States.
Regarding the NDSAT satellite MSS image (resolution 80 m), the H-shaped building in the MSS image has disappeared as shown in Figure 7 (C), so the corresponding Fourier descriptor is shown in Figure 8 (0 ), it is sufficient to use up to N=4.

The present invention is applicable not only to the field of moat sensing as shown in the above embodiments, but also to industrial inspection, medical image processing, and the like.

For example, a suitable application target of the first embodiment is automation of inspection for scratches on products in industrial inspection. By selecting a set of photographed images of the product, candidate points for flaws are narrowed down, and the shape of only the candidate points is precisely analyzed to extract flaws.

Furthermore, the second embodiment may be applied to a system that uses the same detector to inspect for scratches on the inside or outside of an object. In this case, assuming that scratches have a specific shape but different sizes, when extracting scratches of a specific size from a photographed image, one set of training data is set regardless of the size of the scratch. You should do it.

〔Effect of the invention〕

As described above, according to the present invention, there is provided an image binarization processing unit, a shaped object trace amount calculation processing unit that calculates a shape feature amount from the binarized image, a reference shape feature amount file, and the shaped object trace amount. In an image analysis system having a shape feature comparison detection processing unit that compares the shape feature calculated by the calculation processing unit and the reference shape feature in the reference shape feature file,
The shaped object trace amount calculation processing section is configured from a coarse shape feature amount calculation processing section and a precise shape feature amount calculation processing section,
The shape feature comparison and detection processing section includes a coarse shape feature comparison and detection processing section and a fine shape feature comparison and detection processing section, and the rough shape is extracted from the image binarized in the binarization processing section. After the candidate object is selected by the feature calculation processing section, the precise shape feature comparison detection processing section compares only the candidate object with the reference shape feature in the reference shape feature file. This has the remarkable effect of realizing an image processing method that can improve the accuracy of extracting minute objects in images without increasing processing time.

[Brief explanation of drawings]

Fig. 1 is a flowchart of image shape feature extraction processing which is an embodiment of the present invention, Fig. 2 is a diagram showing a configuration example of an image analysis system which is an embodiment of the present invention, and Fig. 3(a). ~
(c) is a diagram showing a processed image, FIG. 4 is an explanatory diagram of a Fourier descriptor, FIG. 5 is a diagram showing an image processing system that is another embodiment of the present invention, and FIG. 7(a) to (Q) are explanatory diagrams of model image shapes and sensor photographed images, and FIGS. 8(a) to (c) are explanatory diagrams of their Fourier descriptor tables. 11: Input satellite image, 13: Shape parameter calculation process for group selection, 14: Target candidate detection process, 16: Image enlargement process in enlarged area, 18: Shape parameter calculation process for precise selection, 19: Target candidate detection process, zl ;Satellite image file,
22.22A: Image analysis processing unit, 23: Reference shape parameter file for group selection, 24: Reference shape parameter file for precision selection, 25: Image display device, 26: Model figure Fourier descriptor file. Kujinsuke ■ Small month 1 win, J and 7th eyes (moan (b) (C) Fig. 1

Claims (1)

[Claims] 1. An image binarization processing unit, a shape feature calculation processing unit that calculates a shape feature from the binarized image, a reference shape feature file, and the shape feature calculation processing unit In an image analysis system having a shape feature comparison detection processing unit that compares the shape feature calculated in the above with the reference shape feature in the reference shape feature file, the shape feature calculation processing unit includes:
The shape feature comparison and detection processing section is composed of a coarse shape feature comparison and detection processing section and a precise shape feature comparison and detection processing section. After selecting a candidate object by the rough shape feature calculation processing section from the image binarized by the binarization processing section, the fine shape feature comparison detection process is performed only for the candidate object. An image feature extracting processing method, characterized in that a comparison is made with a reference shape feature in the reference shape feature file by means of the reference shape feature file. 2. The process of selecting a candidate object by the rough shape feature amount calculation processing unit is a process of selecting an object having the maximum area within a frame surrounding the candidate object. Image feature extraction method described in Section 1. 3. An image binarization processor, a shape feature calculation processor that calculates shape features from the binarized image, a reference shape feature file, and the shape feature calculated by the shape feature calculation processor. In an image analysis system having a shape feature comparison detection processing unit that compares the shape feature with a reference shape feature in the reference shape feature file, the shape feature is used as a Fourier descriptor, and A Fourier descriptor calculation processing unit for a model figure for training data to be accumulated, a file for storing the training data, and a Fourier descriptor selection processing unit for selecting Fourier descriptor information matching an image resolution from the file are provided. An image feature amount extraction processing method, characterized in that the Fourier descriptor selection processing unit selects the Fourier descriptors of a number of terms corresponding to the resolution of the input image.