JPH06333051A - Image processor - Google Patents

Abstract

PURPOSE:To provide an image processor which has the satisfactory recognizing capability to such an image that is rotated, magnified or reduced, etc. CONSTITUTION:An image processor is provided with an edge detecting means which detects the edges of the input images, an image input means 305 which irradiates a coherent beam 300 and inputs an edge image detected by the edge detecting means to a system, a computer hologram 306 which superposes the phase information on the information supplied to the system, a Fourier transformation optical system 307 which applies the Fourier transformation on the phase information, and a detecting means 309 which detects the information supplied to a logarithmic pole coordinate transformation surface 308, i.e., a Fourier transformation surface. Then the edge of an input image is detected and this image is defined as an input image of a logarithmic pole coordinate transformation optical system. Thus it is possible to surely detect the magnification/reduction value, to improve the recognizing capability of the rotated and magnified/reduced images to a reference image, and to improve the detecting capability of the magnification/reduction value respectively.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image processing apparatus,
In particular, the present invention relates to an image processing device capable of recognizing an image that has been deformed by rotation, enlargement / reduction, or the like.

[0002]

2. Description of the Related Art There is a strong need for an image recognition device such as a visual device for an industrial robot or a product inspection device for an automation line. Conventionally, there has been a strong demand for a recognition device using a correlator using an optical matched filter. Development has been done vigorously. Further, recently, development of a recognition device using a neural network has been actively conducted. However, all of these recognition devices have a low ability to recognize an image that has been deformed by shifting, rotating, scaling, or the like, that is, the generalization ability is low, and improvement of this generalization ability is not practical. Had been the biggest challenge of.

One answer to this problem is David Casasent of Carnegie Mellon University (David C
asasent et.al. "Real-time deformation invariant opt
icalpattern recognition using coordinate tranforma
tions ", Appl. Opt., Vol.26, pp.938-942 (1987)). The method of using their correlators is as follows. First, as shown in FIG. 101 is irradiated with coherent light 102, and a computer-generated hologram (CGH) 103 is irradiated on the light 101.
When the Fourier transform lens L ₁ 104 performs the Fourier transform after superimposing the phase information for the coordinate transform by, the information obtained by performing the desired coordinate transform on the input image is obtained on the coordinate transform surface 105. In the case of the above-mentioned document, for example, the desired coordinate transformation is a logarithmic polar coordinate transformation in which even if transformations of scaling and rotation occur in the input image, those transformations are transformed into shift amounts. Here, in the logarithmic polar coordinate transformation, the input image f (x, y) is converted into X = ln (x ² + y ² ) ^1/2 , Y = −arctan (y
/ X) gives an image converted into two pieces of information of distance information and angle information.

Next, as shown in FIG. 10, the information after the logarithmic polar coordinate conversion is applied to a liquid crystal television (LCTV).
a double diffraction system consisting of two Fourier transform lenses L ₂ 107 and L ₃ 108,
Matched spatial filter (MSF) created by applying the same logarithmic polar coordinate transformation to the reference image
The coherent light 110 is incident on the shift invariant correlation optical system in which the filtering by 109 is combined, and is incident. The camera 111 detects the correlation peaks of the reference image and the test image on the correlation plane P, and recognizes them. Is to do. In fact, Casasent et al. In the above-referenced document show an example in which good matching results are obtained for rotation and scaling deformations using the above method.

[0005]

[Problems to be Solved by the Invention] However, Casase
In the method of nt Mr. like, the conversion of the angular direction is performed properly, the conversion of log _e r direction is not sufficient. For example,
When a logarithmic polar coordinate transformation is performed on a square input image as shown in FIG. 3A by simulation, the image becomes as shown in FIG. 3B. This is because the input image is converted to r to log _er. compared the log _e r direction information are converted into extremely narrow area, the feature quantity in this direction would be longer difficult extraction in the image to be detected. Further, the logarithmic polar coordinate transformation of the circular input image of FIG. 7A is as shown in FIG. 7B, but the difference is not so remarkable as compared with FIG. 3B, and especially, they show. In an actual optical system, the difference becomes more difficult to distinguish due to an error in manufacturing CGH, influence of diffraction, and the like. Therefore, even if a correlation optical system using MSF filtering as described above is combined, the recognition rate actually becomes low.

Further, when the image of FIG. 4 (A) obtained by enlarging the square of FIG. 3 (A) is used as the input image, the image of the logarithmic polar coordinate conversion is as shown in FIG. 4 (B). The intensity distribution in the vicinity of the alternate long and short dash line in FIG. 3 (B) is shown in FIG. 3 (C), and the intensity distribution in the vicinity of the alternate long and short dash line in FIG. 4 (B) is shown in FIG. 4 (C). Here, in order to recognize that the image of FIG. 4A that has undergone the enlargement deformation is the same object as the image of FIG. 3A, the image of FIG. 3B must be equal to the image obtained by shifting the position of the image of FIG. 3B. However, each intensity distribution map 4
As can be seen by comparing (C) and FIG. 3 (C), FIG.
In the case of (C), the light intensity is much higher, and when the same threshold value is set and the images are binarized and compared, the log _e r of both images is compared.
It is clear that the widths of the directions are so different that the conditions are not satisfied. In this case, in order to satisfy this condition, it is easy to change the threshold value according to the size of the input image or adjust the intensity according to the r direction of the input image. However, it is not practical because it has a complicated mechanism and causes a large loss of light quantity.

As described above, the log of the input image
_{In order} to express the information in the er direction conspicuously, and to use the amount of image scaling as the amount of positional deviation and to accurately recognize images that have undergone scaling deformation, the method of Casasent et al. Practically difficult.

The present invention has been made in view of the above circumstances, and an object thereof is to provide an image processing apparatus having sufficient recognition ability for an image that has been deformed such as rotated or enlarged or reduced. Is.

[0009]

In order to solve the above problems and achieve the above objects, an image processing apparatus according to the present invention comprises edge detection means for detecting the edges of an input image, and substantially parallel coherent light. Image input means for irradiating coherent light to input the edge image detected by the edge detection means into the system, and phase for logarithmic polar coordinate conversion to image information input into the system by this image input means A computer generated hologram for superimposing information, a Fourier transform optical system for Fourier transforming image information on which phase information is superimposed, and a logarithm output on a logarithmic polar coordinate conversion surface that is the Fourier transform surface of this Fourier transform optical system. And a detection unit for detecting polar coordinate conversion information.

[0010]

In the present invention, for example, if the edges of the square input image of FIG. 3 (A) are taken, the image of FIG. 5 (A) is obtained. If logarithmic polar coordinate conversion is performed on this image, the result of FIG.
It becomes the image of (B). On the other hand, similarly, the edge image of the circular input image of FIG. 7A is as shown in FIG. 8A, and if this is subjected to logarithmic polar coordinate conversion, it becomes the image of FIG. 8B.
The difference between the images in FIG. 5B and FIG. 8B is that FIG.
This is more remarkable than the difference between (B) and FIG. 7 (B). Furthermore, even if the optical system is affected by aberration, diffraction, etc., the difference can be clearly recognized. Therefore, the feature amount can be detected even if it cannot be actually detected, and the recognition rate is improved when the feature amount is recognized by the above-mentioned correlation optical system. Further, an image in which the edges in FIG. 4A in which the square in FIG. 3A is enlarged is detected is shown in FIG. 6A, and an image obtained by logarithmic polar coordinate conversion of this is shown in FIG. 6B. This FIG. 6 (B) is
It can be seen that the image in FIG. 5B is almost equal to the image displaced by the amount corresponding to the enlargement amount. In this way, by taking the edge of the input image and then performing logarithmic polar coordinate transformation, the image that has undergone scaling deformation is
It becomes possible to convert the reference image as an image that is accurately displaced by the amount of enlargement / reduction, and if this converted image is input to the above-mentioned correlation optical system, more accurate recognition is possible than in the conventional example. It is clear that further,
Since the position of the detected correlation peak accurately corresponds to the amount of change in the scaling of the input image, it is possible to more accurately and instantly determine how much the scaling is expanding or shrinking with respect to the reference image. I understand.

[0011]

EXAMPLES Examples 1 and 2 of the image processing apparatus of the present invention will be described below.
Will be described with reference to the drawings. [Embodiment 1] Embodiment 1 of the present invention is an image processing apparatus for recognizing an input image in real time by the optical system shown in FIG. This device is roughly composed of three optical systems: an edge detection optical system, a logarithmic polar coordinate conversion optical system, and a correlation optical system.

In the edge detection optical system, the image to be recognized is displayed on the display device 201, such as a liquid crystal display device. The displayed image is irradiated with coherent light 300, read by the transmitted light, and further optically written in the spatial light modulator 202. In the spatial light modulator 202, the applied voltage is a pulse voltage, and the applied voltage is controlled, whereby the edge image of the written image can be displayed on the reading light side. Therefore, the displayed edge image is irradiated with the reading light 204 through the beam splitter 203, and the edge image is read by the reflected light. The edge image is detected by the image reading device 304 via the image plane 205.

Next, the edge image detected by the image reading device 304 is used as an image input means 305 such as liquid crystal.
And the input image of the log-polar coordinate conversion optical system.
The configuration of the optical system is the same as that in the conventional example log-polar optical system of FIG. 9, is irradiated with coherent light 300 on the displayed image is subjected to log-polar transformation CGH306, further Fourier transform lens having a focal length f ₁ by going through the L ₁ 307, obtain the log _e r-theta-converted image to the coordinate conversion surface 308 located on the back focal plane of the Fourier transform lens L ₁ 307. At this time, the resulting image is, as described above, is remarkably displayed image the log _e r direction of features of the input image, also, when the image scaling of an image, the scaling amount It is an image that is misaligned by an amount. This image is detected by detection means 309 such as a CCD.

The image detected by the detecting means 309 is
Further, it is displayed on a display device 310 such as a liquid crystal and used as an input image of the correlation optical system. The correlation optical system is substantially the same as the correlation optical system of the conventional example shown in FIG. 10. The display device 310, the substantially parallel coherent light 300, and the coordinate conversion information input into this system by these are Fourier-transformed. Focal length f
₂ Fourier transform lens L ₂ 311 and MSF 312 for superimposing reference image information for recognition placed on the rear focus of this Fourier transform lens L ₂ 311;
Another Fourier transform lens L ₃ 313 having a focal length f ₃ for Fourier transforming the superimposed information, and a correlation surface 314 placed at the rear focus of the other Fourier transform lens L ₃ 313. The image reading device 315 for detecting the correlation peak output to the above.

First, the MSF 31 of the image serving as a reference for recognition.
The manufacturing method of No. 2 will be described. This reference image is
It is displayed on the display device 201 of the edge detection optical system. The displayed image is subjected to image processing of edge detection and logarithmic polar coordinate conversion by passing through an edge detection optical system and a logarithmic polar coordinate conversion optical system, and the image is input to the correlation optical system. The image input to this correlation optical system is further Fourier-transformed by the Fourier transform lens L ₂ 311 to obtain the MSF 3
The Fourier transform image is made at the position of 12, and the hologram dry plate is placed at this position, this information is recorded, and MSF31 is recorded.
Set to 2. The MSF 312 may of course be formed as a computer hologram by calculation.

Next, the recognition process will be described. First, the image to be recognized is displayed on the display device 201. The image input to the correlation optical system is the Fourier transform lens L ₂ 31.
Fourier transform by 1, and this information is
2 is superposed on the information, and finally, the Fourier transform is performed again by the Fourier transform lens L ₃ 313 to obtain the correlation surface 314.
The peak of the correlation between the reference image and the recognized image is output. Since the height of this correlation peak indicates the degree of coincidence of these images, the ones having a high peak may be judged to be the same. In this correlation optical system, the MS in the system
Since the position of F312 is exactly the Fourier transform plane of the input image to this correlation optical system, even if the input image to the correlation optical system is shifted, the same image information is obtained at this MSF312 position. . Therefore, even if the input image to the logarithmic polar coordinate conversion optical system is an enlarged or reduced image of the reference image, an image displaced by an amount corresponding to the amount of enlargement or reduction can be obtained. Since the accuracy is similar to, even if the rotation / enlargement / reduction deformation is present in the input image, the correlation peak of this correlation optical system is the same as the correlation peak between the reference images without deformation.

From the above, it can be seen that the image processing apparatus of the present embodiment recognizes that rotation / enlargement / reduction is permitted by the edge detection optical system, the logarithmic polar coordinate conversion optical system, and the correlation optical system. As described above, by using the edge image of the input image as the input image of the logarithmic polar coordinate conversion optical system, the input image to the correlation optical system is the log of the recognition target image.
It represents a feature value of _e r direction significantly, which is an image in which misaligned by an amount corresponding to the scale of the image. Therefore, the recognition by the correlation peak obtained by the correlation optical system can surely represent the enlargement / reduction amount of the image to be recognized by the positional shift amount of the correlation peak without lowering the recognition rate.

Although the image input means is a liquid crystal display device here, it can be realized by using other spatial modulation elements.

Further, in this embodiment, as the edge detecting means, the method of controlling the voltage of the spatial light modulator has been shown. However, an optical system similar to the above-mentioned correlation optical system is used, and the space is used instead of the MSF. An edge image of the input image can be obtained by performing filtering using a filter that cuts the 0th-order light portion in the frequency domain. Further, as an edge detecting means, if the high speed is sacrificed to some extent, the input image is C
Various other modifications are conceivable. For example, it can be realized by using a means for detecting an edge by capturing an image captured by an image sensor such as a CD into a computer with a frame memory or the like and taking a Laplacian of the image on the computer. .

[Embodiment 2] This embodiment is an apparatus for recognizing an input image (pattern classification) in real time by the optical system shown in FIG. The edge detection optical system and the logarithmic polar coordinate conversion optical system in the figure are composed of exactly the same elements as in the first embodiment. Therefore, an image obtained by further Fourier transforming the image subjected to edge detection / logarithmic polar coordinate conversion is detected by the image reading device 413 such as a CCD. Here, the detection means 413
The image detected by 1. is the same as that described above. Even if the image input to the display device 201 is the reference image rotated / displaced / enlarged / reduced, the image input by the display device 410 Is subjected to Fourier transform by the Fourier transform lens L ₄ 411 having the focal length f ₄ , so that the same image as that when the reference image is input is formed on the free Fourier transform surface 412 and detected by the image reading device 413. This image is input to the neural network 414.

In this embodiment, as the neural network 414, a backpropagation model is constructed by software on a computer. In this back-propagation model, it is necessary to first determine the synaptic weight between the layers of the neural network 414 by learning. This is performed by sequentially inputting images to be recognized (pattern classification) on the display device 201.
Finally, the Fourier transform images of those images are further subjected to logarithmic polar coordinate transformation to the neurons of the input layer of the neural network 414, and the resulting information is sequentially input as image information obtained by Fourier transform. At this time, the output layer of the neural network 414 is output. The back-propagation learning rule may be used to determine the synapse weights between the layers so that the firing of the neurons is selectively performed corresponding to the image to be recognized (pattern classification). At the time of recognition (pattern classification), the synapse load between the layers may be fixed, an image to be recognized (pattern classification) may be presented on the input surface by the display device 201, and firing of neurons in the output layer may be detected. .

With the above-described structure, in addition to the pattern classification ability of a normal neural network, even if the input image is deformed by rotation or enlargement / reduction, by using these optical systems, the image reading device 413 can obtain the same image. Since it is obtained, it is obvious that the improvement of the generalization ability that can be recognized as the same object as the reference image can be realized. Needless to say, the recognition ability is remarkably improved by using the edge image, as in the previous embodiment. Further, in the above embodiment, the back propagation model was considered as the neural network, but it goes without saying that the generalization ability can be similarly improved even when another model such as the Hopfield model is used.

Although the image to be recognized is displayed on the display device 201, the image to be recognized is subjected to Fourier transform,
If the image is displayed on the display device 201, the image displayed on the display device 201 will be the same even if the image to be recognized is displaced. Therefore, it is possible to realize a system that enables recognition even if the image to be recognized is displaced.

[0024]

As is apparent from the above description, the image processing apparatus of the present invention detects the edge of an input image and uses that image as the input image of the next logarithmic polar coordinate conversion optical system to output the output. was significantly characteristic of log _e r direction in the image, since the scaling amount can be reliably detected, the recognition capability of the image obtained by reducing rotation and enlarged relative to the reference image, and scaling the amount of detection capability is improved.

[Brief description of drawings]

FIG. 1 is a diagram showing an optical system of an image processing apparatus according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an optical system of an image processing apparatus according to a second embodiment of the present invention.

FIG. 3 is a diagram showing an example of an input image and a converted image thereof.

FIG. 4 is a diagram showing another input image and its converted image.

5 is a diagram showing an input image obtained by edge detection of the image of FIG. 3 and a converted image thereof.

6 is a diagram showing an input image obtained by edge detection of the image of FIG. 4 and a converted image thereof.

FIG. 7 is a diagram showing another input image and its converted image.

8 is a diagram showing an input image obtained by edge detection of the image of FIG. 7 and a converted image thereof.

FIG. 9 is a diagram showing a conventional logarithmic polar coordinate conversion optical system.

FIG. 10 is a diagram showing a conventional correlation optical system.

[Explanation of symbols]

201 ... Display device 202 ... Spatial light modulation element 203 ... Beam splitter 204 ... Readout light 205 ... Image plane 300 ... Coherent light 304 ... Image reading device 305 ... Image input means 306 ... CGH (Computer hologram) 307 ... Fourier transform lens L ₁ 308 ... coordinate transform plane 309 ... detecting means 310 ... display device 311 ... Fourier transform lens L ₂ 312 ... MSF (matched filter) 313 ... Fourier transform lens L ₃ 314 ... correlation surface 315 ... image reading apparatus 410 ... display device 411 ... Fourier Conversion lens L ₄ 412 ... Frier conversion surface 413 ... Image reading device 414 ... Neural network

Claims (1)

[Claims] 1. An edge detecting means for detecting an edge of an input image, a substantially parallel coherent light, and an image for irradiating the coherent light and inputting the edge image detected by the edge detecting means into a system. Input means, a computer generated hologram for superimposing phase information for logarithmic polar coordinate conversion on the image information input into the system by the image input means, and a Fourier transform for Fourier transforming the image information on which the phase information is superimposed. An image processing apparatus comprising: a conversion optical system; and a detection unit for detecting logarithmic polar coordinate conversion information output to a logarithmic polar coordinate conversion surface which is a Fourier transform surface of the Fourier transform optical system.