JP3074292B2 - Template matching processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for pattern matching between images in image processing.

[0002]

2. Description of the Related Art In recent years, while image processing application techniques have been introduced in various fields, various processing methods for comprehensively analyzing obtained image data and extracting highly accurate and diversified features have been attempted. .

In the field of FA, there are many uses of pattern matching processing for obtaining correspondence between images. In particular, template matching processing for scanning and calculating the similarity between a template image and a texture image to obtain the position of an object is performed by: This is the most common method used for processing such as object positioning, tracking of a moving object, and three-dimensional shape measurement.

Here, the template matching process will be described. As shown in FIG. 5, assuming template image data g (x, y) and texture image data f (x, y) including the target to be obtained, normalized cross-correlation R (x, y)
y) represents the center coordinates of the image data g (x, y) as (k,
ι), it is calculated by the formula. This calculation is performed on the pixels of the entire texture image, and the pixel position at which the normalized cross-correlation value is maximized is obtained.

FIG. 6 shows the distribution of the normalized cross-correlation values calculated by the equations. As a result, the position having the highest correlation can be obtained with original pixel accuracy.

Next, the case where the position of the object is obtained with sub-pixel accuracy will be described. The template image data g (x, y) and the texture image data f shown above
A density interpolation process is performed on (x, y) to generate grayscale image data on the order of sub-pixels, and a similar operation may be performed on the entire interpolation pixel using an equation. Here, an example of the density interpolation processing will be described with reference to FIG. The position of the original pixel image data is represented by u, v coordinates, and the coordinate position is (u, v),
(U, v + 1), (u + 1, v), (u + 1, v + 1)
And the density values are f (u, v), f (u, v + 1),
f (u + 1, v) and f (u + 1, v + 1), and the density at the position of (u ₀ , v ₀ ) after density interpolation is f (u ₀ , v ₀ ).
Then, the density value f (u ₀ , v ₀ ) is calculated by an equation. f (u ₀ , v ₀ ) = f (u, v) (1−α) (1−β) + f (u + 1, v) α (1−β) + f (u, v + 1) (1−α) β + f ( u + 1, v + 1) αβ ...

Α and β are, for example, 0.1 and 0.2 to 1 as shown in FIG. 8 when the original image data is interpolated by dividing the original image data into ten equal parts (10 × 10 = 100 times interpolation). To obtain the density value at each point. As a result, the position having the highest correlation can be obtained with sub-pixel accuracy. FIG. 9 shows an example of a conventional processing flow for obtaining a corresponding point with sub-pixel accuracy.

[0008]

As described above, a large amount of solid image data is used in the template matching process.
When trying to find the corresponding position with sub-pixel accuracy, the data amount of the original pixel is enlarged by the interpolation magnification (for example, 10 × 10 = 100 times with 0.1-pixel accuracy).
In addition, since the correspondence scanning is performed at all points of the texture image, a problem that a considerably long calculation time is required to find a place that best matches the template image pattern occurs.

As solutions to these problems, it is conceivable to introduce a CPU that operates at a very high speed or to deal with hardware such as a parallel distributed processing configuration using a plurality of MPUs. However, this leads to an increase in the cost and size of the processing device. Not a good idea. On the other hand, as a software solution, a method of performing the cross-correlation calculation using the fast Fourier transform, or when the partial sum of the cross-correlation calculation exceeds a certain threshold, the calculation is interrupted to improve the calculation efficiency. SSDA method (Sequen
tial Similarity Detection
Algorithm) or a two-step coarse / fine search method that narrows the corresponding range by interlaced scanning and then scans the entire range is proposed. However, the only way to increase the position detection accuracy is to increase the density interpolation magnification. Along with this, there is a problem that the image data also increases, and the calculation time eventually increases.

The present invention has been made in view of the above circumstances,
An object of the present invention is to provide a template matching processing method for finding a corresponding position with sub-pixel accuracy in a very short time without performing a density interpolation process in a template matching process for finding a corresponding position haze between images.

[0011]

A maximum cross-correlation value based on original pixel data and a pixel position around the maximum cross-correlation value element position, which are obtained by associating scanning between grayscale images picked up by a television camera. The corresponding position of sub-pixel accuracy is obtained by using the difference or the slope of the cross-correlation value of.

[0012]

The difference between the maximum cross-correlation value obtained by the normalized cross-correlation calculation based on the original pixel data and the cross-correlation value between the pixel positions surrounding the maximum cross-correlation pixel position or the inclination is calculated. Since it is possible to obtain corresponding points of pixel accuracy, it is not necessary to perform template matching processing on large image data on which density interpolation processing has been performed as in the related art.

[0013]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 illustrates a template matching processing flow according to the present embodiment, which will be described in the order of the processing flow. First, an analog video signal from a television camera is fetched and sequentially A / D converted. Next, as shown in FIG. 5, an area of the template image data g (x, y) including the object and the area of the texture image f (x, y) are set on the digitized captured image data. Next, the normalized cross-correlation calculation will be described. Assuming that the center coordinates of the image data g (x, y) are (k, ι), the normalized cross-correlation R (x,
y) is calculated by an equation. This calculation is performed on the pixels of the entire texture image, and the pixel position at which the normalized cross-correlation value is maximized is obtained. FIG. 2 shows a three-dimensional distribution model of the normalized cross-correlation value calculated by the equation. The x and y axes are
The coordinate axis of the pixel and the z-axis represent the cross-correlation value at each pixel position. In the drawing, W _{1 to} W ₉ are cross-correlation values at each pixel position, and the degree of correlation is represented by a height difference. Highest point of the cross-correlation value is a pixel having a value of W _2, so that corresponding points of the sub-pixel accuracy exists in this pixel.

Next, a description will be given of a method of obtaining a corresponding point of sub-pixel accuracy. Figure 3 is a cross-correlation value in FIG. 2, x-axis of囲Ri of pixels having a maximum cross-correlation value W _2, y-axis direction 4
In this model diagram, points corresponding to sub-pixel accuracy are obtained from this state. First, a method for finding a corresponding point of sub-pixel accuracy in the x-axis direction will be described with reference to FIG. W ₂ is the maximum cross-correlation value, W ₁ and W ₃ are the cross-correlation values of both neighboring pixels, and θ ₁ is a line segment connecting W ₁ and W ₂ , and a line segment vertically lowered from W ₁ to OW _2. Angle with,
theta ₂ is the angle between the line segment drawn vertically OW ₂ the line segment and _{W 3} connecting _{W 3} and _{W 2.} Since the corresponding point of the sub-pixel accuracy is that OW ₂ is at a position where θ ₁ = θ ₂ is satisfied (shown by a dotted line in the figure), it can be obtained by the equation. t
than anθ ₁ = tanθ ₂

Similarly, the corresponding point of the sub-pixel accuracy in the y-axis direction can be obtained by an equation using W2, W4, and W5.

From the above results, it is required that the corresponding point of sub-pixel accuracy exists at a position shifted from the pixel center by (Δι _x , Δι _y ). If the difference between the maximum cross-correlation value W _{2 of the} original pixel and any of the cross-correlation values W ₁ , W ₃ , W ₄ , and W ₅ at pixel positions around the pixel position of the maximum cross-correlation value is minute (a certain constant or less) , Or zero)
The same calculation is performed using the maximum cross-correlation value pixel and the cross-correlation value of a pixel position around a region including a pixel having a small cross-correlation value with a small difference from the maximum cross-correlation value in consideration of a quantization error and the like. Should be implemented. Also, as is clear from the calculation formula,
The position detection accuracy includes the maximum cross-correlation value W ₂ and the cross-correlation values W ₁ and W of pixel positions around the maximum cross-correlation pixel position.
If there is a difference from ₃ , W ₄ , and W ₅ , or a gradient, it can be increased as much as possible without taking into account quantization error and the like.

[0017]

As described above, the present invention utilizes the cross-correlation value obtained from the data of the original pixel, so that it is not necessary to handle large image data subjected to density interpolation processing as in the prior art, and it is extremely fast. In addition, there is a great effect that high-precision matching between both images can be realized with a low-cost and compact device.

[Brief description of the drawings]

FIG. 1 is a processing flow of an embodiment.

FIG. 2 is a three-dimensional model diagram of a cross-correlation distribution.

FIG. 3 is a model diagram for obtaining a corresponding position of sub-pixel accuracy.

FIG. 4 is a diagram for explaining a method of obtaining a corresponding position in the x-axis direction.

FIG. 5 is a diagram for explaining a cross-correlation.

FIG. 6 is a diagram showing distribution of cross-correlation.

FIG. 7 is a diagram for explaining density interpolation;

FIG. 8 is a diagram for explaining density interpolation data;

FIG. 9 is a processing flow for obtaining a corresponding point with conventional sub-pixel accuracy.

──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Mitsutoshi Akai 3-11-5 Iwamotocho, Chiyoda-ku, Tokyo Examiner, Daido Electric Industry Co., Ltd. Futoshi Yasuda (56) References JP A) JP-A-7-43309 (JP, A) JP-A-4-194702 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T 7/00

Claims (1)

(57) [Claims] 1. An image processing apparatus for digitizing an analog video signal from a television camera and performing measurement and analysis processing by means of an arithmetic element, for performing cross-correlation between images in the measurement and analysis processing. In the calculation and the template matching process for finding a corresponding position from the maximum cross-correlation value of the cross-correlation calculation result, the maximum cross-correlation value based on the original pixel data, obtained by the cross-correlation calculation, and the maximum cross-correlation value pixel A template matching processing method characterized in that a corresponding position with sub-pixel accuracy is determined by using a difference or a slope of a position with a cross-correlation value of peripheral pixels.