JPH05297325A - Artificial retina and artificial visual sensation device - Google Patents

Abstract

PURPOSE:To obtain the artificial retina necessary for easily recognizing plural objects to be recognized at a high speed from images contg. these objects and the artificial visual sensation device constituted by using this artificial retina. CONSTITUTION:This device has an imaging means 2, an artificial eyeball 3 which has the artificial retina 1 consisting of first artificial retina cells for detecting a boundary between brightness and darkness by optical filtering disposed in a central visual field 1a and second artificial retina cells for detecting an object position by optical filtering disposed in a peripheral visual field 1b, a neural network 4 which recognizes the patterns of the object by receiving the information detected by the first artificial retina cells, a means which determines the object to be recognized next from the information detected by the second artificial retina cells of the artificial retina 1 and an artificial eyeball moving means 5 which moves the artificial eyeball 3 toward this object to be recognized. The specific object to be recognized is selected easily at a high speed from the images including the plural objects to be recognized, by which the specific object is easily recognized.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an artificial retina and an artificial vision device effective for recognizing a plurality of objects to be recognized from an image.

[0002]

2. Description of the Related Art The scenes we see every day, so-called natural images, contain many target objects to be recognized,
We recognize them unconsciously and think,
Making decisions and acting. Then, in the recognition process, the position where the contrast suddenly changes in the image by the initial visual mechanism, that is, the boundary part of the light and dark is detected, this information is sent to the brain in the subsequent stage, and the object is recognized. It is said. This initial visual mechanism, as shown in FIG. 10, includes photoreceptor cells 111, horizontal cells 112, and bipolar cells 11
It is composed of 3. The mechanism will be briefly described. First, the photoreceptor cells 111 receive light from the image, and only the photoreceptor cells 111 in the portion receiving the light are excited. This excitement is transmitted to the horizontal cells 112 in the lower layer.
12 is planarly connected to about 6 adjacent horizontal cells 112, and due to the transmission through this connection, the excitation of the surrounding horizontal cells 112 is gently changed so as to be transmitted through the network. Furthermore, the bipolar cell 113 is the photoreceptor cell 11
1, and is connected to two horizontal cells 112, and excites corresponding to the difference in excitement intensity between the two cells 111 and 112. This excitement is further transmitted to the lower layers, and eventually reaches the brain,
Object recognition and feature extraction are performed.

Actually, considering a boundary portion of light and darkness such as a contour of an object as an image, in this initial visual mechanism, first,
As shown by the curve A in FIG. 11A, the excitement intensity of the photoreceptor cell 111 changes rapidly. On the other hand, the excitation intensity of the horizontal cell 112 changes gently as shown by the curve B. Therefore, in the bipolar cell 113, these two cells 111,
Corresponding to the difference in the excitement intensity of 112, the curve C is shown in FIG.
As shown in, just the light-dark boundary is excited, and the information on the light-dark boundary is transmitted to the brain, where the recognition process is performed.

The initial visual mechanism having the function of detecting the bright and dark boundaries was concretely constructed by an electric circuit and was prototyped as a chip by C. of California Institute of Technology. A. Mead et al. (CA Mead and MA Maho)
wald, “A Silicon Model of of
Early Visual Processing ”,
Neural Networks, vol. 1, pp.
91-97 (1988)). Their element is shown in FIG.
As shown in a partially enlarged circuit diagram, photoreceptor cells are constituted by photodiodes 201, horizontal cells are constituted by six resistance networks 202 connecting the centers and apexes of hexagons, and bipolar cells are constituted by amps 203. In the bright and dark parts of the image, the voltage of the photodiode 201 changes stepwise. On the other hand, the voltage of the node of the resistance network 202 changes gently. By taking this voltage difference, the voltage of the amplifier 203 changes only at the bright / dark boundary portion. The change in voltage of these elements is similar to the change in the excitation intensity of each cell shown in FIG. 11, and it is possible to detect the light-dark boundary, that is, the contour of the object, as in the case of human initial vision.

Further, although not directly conscious of the initial visual mechanism, a method of optically detecting a light-dark boundary is described in K. K. G.
Proposed by Mr. Birch (K. G. Bir
ch, “A Spatial Frequency F
ilter to Remove Zero Free
quency ”, OPTICA ACTA, vol.1
5, No. 2, pp. 113-127 (1968)).

This method is a method of detecting a light-dark boundary in an image by cutting the 0th-order light portion in the spatial frequency domain. The interpretation of their method in consideration of the initial visual mechanism is as follows. For simplicity, consider only the one-dimensional direction in FIG. The brightness information in the image is φ
(X) (FIG. 13 (a)), Fourier transform this,
Filter that cuts 0th order part {1-rect (ξ /
b)}, it can be expressed as F {φ (x)} · {1-rect (ξ / b)} (1) (FIG. 13 (b)). However, F {} represents a Fourier transform, and rect (x) is And b of rect (ξ / b) is determined such that {1-rect (ξ / b)} almost cuts the 0th-order peak of the diffraction pattern. Furthermore, when the equation (1) is subjected to inverse Fourier transform and is set as f (x), f (x) is f (x) = φ (x) −φ (x) * sinc (x / b)・ ・ It can be expressed as (2). However, * represents convolution, s
inc (x) is sinc (x) = sin (x / 2) / (x / 2). This f (x) is as shown in FIG.
It is obtained by subtracting φ (x) * sinc (x / b) which is blunted by convolution from the light / dark information φ (x). By comparing this with FIG. 11, φ (x) of the first term of Expression (2) is obtained. x) is similar to the response of photoreceptor cells, φ (x) * sinc (x / b) of the second term of formula (2) is similar to the response of horizontal cells, and f (x) is the bipolar cell. It can be seen that it is similar to the response of. From the above,
It can be said that their method is one of the expression methods of the human early visual mechanism. In addition, since the optical processing actually deals with the intensity, the processed image is represented as | f (x) | ² as shown in FIG. The bright and dark parts are obtained as the sandwiched dark lines.

Therefore, the above-mentioned Mr. Mead and Birch
By detecting the information of the light and dark boundaries of the object with the element and method proposed by him and inputting this to a so-called neural network that mimics the information processing function of the human brain and processing it, the object recognition mechanism is constructed. It's natural to think we can, and this is something everyone can think of.

[0008]

The element and method for detecting the light-dark boundary according to the above-mentioned conventional example, and the recognition mechanism using them, make good use of the human initial visual mechanism, and certainly the object in the image. However, it is difficult to recognize an image including a plurality of recognition target objects. This is because, if there is light-dark boundary information of a plurality of objects in the input image, it must be recognized after separating them into information for each object. Recognition when the number of recognition target objects is limited to one can be realized by simple associative memory on a neural network,
This separation of information for each object requires fairly sophisticated and complicated processing, and its realization is not easy, no matter how much it can be called a neural network. Therefore, some new device is needed.

The elements and methods for detecting the bright and dark boundaries shown in the conventional example also have the following drawbacks. The former Mead et al.'S device is suitable for integration and miniaturization by utilizing the current LSI manufacturing technology,
Electrical delay of the components is unavoidable and the processing speed is relatively slow. In addition, the latter method of Mr. Birch is an excellent method that takes advantage of the high speed of light generation and parallelism due to the optical configuration, but in reality, it is difficult to align the components and the temperature is high. However, it is hard to say that it can be easily realized due to low reliability against vibration and the like.

The present invention has been made in view of the above circumstances, and an object thereof is to use an artificial retina necessary for recognizing an object including a plurality of objects to be recognized easily and at high speed. To provide an artificial vision device that has been used.

[0011]

As described above, when recognizing an object included in an image, it is appropriate to apply the information processing principle of human initial vision and use the light-dark boundary information. Since it is clear that, also in the case of the present invention, this light-dark boundary information is used for the recognition processing. Specifically, first, the following artificial retinal cells are used. This is, as shown in FIG. 1, a Fourier transform lens 10 for performing Fourier transform of information imaged on the input surface 9 which is one lens end surface to a substantially flat end surface on the opposite side, and the input surface 9 of the Fourier transform lens 10. Of the filter 11, which is formed in close contact with the substantially flat end surface on the side opposite to the end surface, and one of the substantially flat end surfaces is adhered to the filter 11, and the filtered information is subjected to inverse Fourier transform, and the other end surface is formed. It is composed of an inverse Fourier transform lens 13 which forms an image again on a certain output surface 12 as real space information.

If a spatial frequency filter that cuts the 0th order light of the Fourier-transformed diffracted light is used as the filter 11 in the construction of such artificial retinal cells, B of the conventional example is used.
The same filtering method as Mr. irch's method,
The bright and dark boundary detection necessary for recognition of objects in an image can be performed in parallel and at high speed. Further, in the case of the present invention, as the Fourier transform lens 10 and the inverse Fourier transform lens 13, lenses having a substantially flat end face, such as a gradient index lens, are used. Therefore, each configuration based on the lens end face is used. In addition to enabling alignment between elements, there is no need for alignment in the tilt and optical axis directions. Further, the 0th-order light cut portion 11a of the filter 11 is
A photosensitizer such as a photoresist or a silver salt emulsion is applied to a substantially flat end surface of the Fourier transform lens 10 opposite to the end surface of the input surface 9 to irradiate the input surface 9 with an actual light flux as a parallel light flux, and 0 Utilizing the fact that the energy of the next light portion is extremely high compared to other portions, if it is directly formed by exposing only that portion, it directly affects the detection of the bright and dark portions,
It is not necessary to perform alignment in the direction perpendicular to the optical axis of this portion 11a having the highest alignment accuracy (in the ξ axis and η axis directions in FIG. 1).

As described above, the structure of the artificial retinal cells described above solves the problems of the alignment accuracy, the reliability of temperature and vibration, etc., which the conventional method of Mr. Birch had, and the reliability of the object. As a preprocessing function for recognition, it can be easily obtained artificial retinal cells that can detect bright and dark boundaries with sufficient accuracy.

Further, in the artificial retina according to the present invention, as shown in FIG. 2, a plurality of the above artificial retinal cells are arranged in parallel, and an image is formed on the output surface 12 (FIG. 1) behind the artificial retinal cells. The detection means 8 for detecting the generated information is provided and configured, but as described above, the light and dark boundary information is obtained by using the 0th-order light cut spatial frequency filter as a filter of all artificial retinal cells. However, since it becomes difficult to easily distinguish and recognize an image including a plurality of recognition target objects, the following measures are taken.

That is, the vicinity of the center of the artificial retina is defined as the central visual field 1a, the artificial retinal cells arranged in the central visual field 1a are defined as the central visual field cells 6, and the filter is constituted by a spatial frequency filter of 0th-order light cut, The bright and dark boundaries of the captured image are detected, and this information is sent to the neural network in the subsequent stage to recognize the object (see FIGS. 4 and 7). The range of the central visual field cells 6, that is, the central visual field 1a is limited to a relatively small area on the artificial retina,
In the area in the image to which it corresponds, almost one object is recognized, so that the recognition function can be realized by a simple neural network as described above.

However, as it is, only one piece of object information that happens to be in the central visual field 1a can be recognized, and the object of the present invention can recognize a plurality of objects in an image including a plurality of objects to be recognized. is not. Here, it is further conceivable that the central visual field 1a is moved to the position of other information to be recognized by some method. For that purpose, it is first necessary to determine the position of the object outside the central visual field 1a. Therefore, in the artificial retina according to the present invention, as shown in FIG. 2, the outside of the central visual field 1a is the peripheral visual field 1b, the artificial retinal cells contained therein are the peripheral visual field cells 7, and the peripheral visual field cells 7 are As shown in FIG. 3, like the central visual field cells 6,
A Fourier transform lens 10 that performs a Fourier transform on the information imaged on the input surface 9 which is one of the lens end surfaces to a substantially flat end surface on the opposite side, and a substantially flat surface on the opposite side of the end surface of the input surface 9 of the Fourier transform lens 10. Filter 1 closely formed on the end face
1. The one end face which is substantially flat is brought into close contact with the filter 11, the filtered information is subjected to the inverse Fourier transform, and the output face 12 which is the other end face is imaged again as real space information from the inverse Fourier transform lens 13. The filter 11 is composed of a high-order light cut spatial frequency filter including a high-order light cut portion 11c, and only low-frequency components in the image, that is, the rough shape of the object is detected,
By using this, the position of the object in the peripheral visual field 1b is detected.

Next, it is necessary to move the central visual field 1a to the next recognition target by using the information from the peripheral visual field cells 7. Therefore, as shown in FIG. 4, an artificial eyeball including an artificial retina 1 as shown in FIG. 2 and an image forming means 2 for transmitting information including an object in an image to be recognized to the artificial retina 1. 3, the artificial eyeball moving means 5 for deciding the position of the object to be recognized next based on the above-mentioned information detected by the peripheral visual field cells 7 and moving the central visual field 1a to the recognition target. Try to move 3.

Summarizing the above with reference again to FIG. 4,
The artificial vision device of the present invention transmits image information including a plurality of recognition target objects onto the artificial retina 1 by the image forming means 2 and takes in almost one object from the image into the central visual field 1a in the artificial retina, The light / dark boundary information is detected, and this information is sent to the neural network 4 in the subsequent stage, where the object is recognized (in the figure, the light / dark boundary of the triangular portion in the image is detected). At the same time, a rough shape of an object other than the currently recognized object is detected in the peripheral visual field 1b in the artificial retina, and the artificial eye moving means 5 further determines the position of the next object to be recognized. When the recognition of the currently recognized object is completed, the central visual field 1a is moved to that position, and the next object is recognized there. In FIG. 4, for example, a square portion is selected as a target to be recognized next, and the central visual field 1a is moved with respect to this image as shown in FIG. Although the above operation is repeated, it behaves just like a human recognizing an object while moving its eyes around.

As is clear from the above description, the artificial retina of the present invention has a Fourier transform lens for forming a Fourier transform image of an input object image on a predetermined surface, and a Fourier transform lens arranged on the predetermined surface. First-type artificial retinal cell consisting of a filter for removing the 0th-order light component of the transformed image and an inverse Fourier transform lens for performing an inverse Fourier transform on the filtered Fourier transform image to return it to real space information, and an input object A Fourier transform lens that forms a Fourier transform image of the image on a predetermined surface, a filter that is arranged on the predetermined surface to remove high-order light components of the Fourier transform image, and an inverse Fourier transform of the filtered Fourier transform image Detecting output light from the second type artificial retinal cells including an inverse Fourier transform lens for returning to real space information and the first and second type artificial retinal cells The element out, is characterized in that it comprises a.

In this case, the Fourier transform lens of the artificial retina cells of the first and second types has an entrance surface and an exit surface which are substantially flat, and a Fourier transform image is formed on the exit surface, and each filter has The inverse Fourier transform lenses of the first and second types of artificial retinal cells are arranged in close contact with the exit surface, and the entrance surface and the exit surface are substantially flat, and the entrance surface is in close contact with each filter. It is desirable that they are arranged so that an inverse Fourier transform image is formed on the exit surface thereof.

Further, the artificial vision device of the present invention is detected by the image forming means for forming the object image, the artificial eyeball provided with the above artificial retina, and the artificial retina cells of the first kind of this artificial retina. A neural network for receiving information to recognize a pattern of an object, a means for determining an object to be recognized next from information detected by the second type artificial retinal cells of the artificial retina, and a target for the recognition And an artificial eyeball moving means for moving the artificial eyeball.

[0022]

In the present invention, the artificial retina is composed of the first type artificial retinal cells capable of recognizing one object information and the second type artificial retinal cells capable of separating a plurality of objects and detecting their positions. , A specific object can be easily and quickly selected from an image including a plurality of recognition target objects and easily recognized.

[0023]

The preferred embodiments of the artificial retina and the artificial vision device of the present invention will be described below. First Embodiment This embodiment relates to the artificial retina in the present invention. As shown in FIG. 1, this is used as the Fourier transform lens 10 and the inverse Fourier transform lens 13 of the central visual field cells for constructing the artificial retina, with a diameter of 1 mm and a pitch length of 1
A gradient index rod lens having a 2.8 mm, a numerical aperture of 0.38, and a central refractive index of 1.557 was used with a length of 0.25 pitch. The one pitch length of the gradient index rod lens is the length at which the image of the entrance end face is formed as an inverted image once at the middle and is formed as an erect image on the exit end face.
The diameter of the substantially flat surface of the Fourier transform lens 10 opposite to the input surface 9 of the gradient index rod lens is about 2 μm.
An aluminum film is formed at a portion of m or less by a method of forming a filter through the above-mentioned actual light ray as a 0th-order light cut portion (11a in FIG. 1) for detecting a light-dark boundary, and further, a diameter of about 50 μm or more. As a high-order light cut portion (11b in FIG. 1; in the case of the present embodiment, about 30th order or more is cut) for removing noise and the like, an aluminum film is formed by a normal mask patterning process, and the filter 11 And Next, this filter 11
The central visual field cell 6 was constructed by bonding the gradient index rod lens of the Fourier transform lens 10 in which is formed and the gradient index rod lens of the inverse Fourier transform lens 13 to each other. Further, as the peripheral visual field cells 7, the same as the central visual field cells 6 except for the filter 11 is used.
As the first example, an aluminum film was formed in a portion having a diameter of about 17 μm or more as a high-order light cut portion (11c in FIG. 3; in the case of this embodiment, cut about 10th order or more) by a normal mask patterning process. As the detecting means 8 (FIG. 2), a ⅔ inch 380,000-pixel CCD (vertical length of about 6.6 mm, horizontal length of about 8.8 mm) is used, and one central visual field cell 6 is provided at approximately the center of the CCD. Peripheral visual field cells 7 in the rest
62 were arranged at a ratio of 7 vertically and 9 horizontally.

In this artificial retina, as an example, FIG. 6 (a)
When an image such as that shown in Fig. 6 (the part shown in black is bright and the part shown in white is dark) is input, as shown in Fig. 6 (b), two pixels are displayed in the central part from the CCD which is the detecting means. The triangle of the double line is obtained (the light and dark boundary part between the double lines),
Image information indicating rough information of other objects is obtained in the peripheral portion.

In this embodiment, the high-order light cut portion 11b is also formed as the filter 11 of the central visual field cell 6, but this is for the purpose of removing noise and the like in the image, and is essentially the 0th-order light portion. Only the cut of 11a is required. Also,
In FIG. 1 to FIG. 3, both end surfaces of the lens used are shown as planes, but this is done to simplify the alignment of the apparatus, that is, the filter 11 of the Fourier transform lens 10 and the inverse Fourier transform lens 13. The side surface needs to be substantially flat, but the opposite surface need not. Further, the input surface 9 and the output surface 12 are also set to the lens end surfaces for the same reason, but this is not necessary if the ease of alignment is sacrificed slightly. Furthermore, it is clear that an artificial retina having the same effect can be obtained by using a graded-index flat plate microlens or the like instead of the graded-index rod lens.

Second Embodiment This embodiment relates to an artificial vision device to which the artificial retina of the first embodiment is applied. The configuration is shown in FIG. 7. In FIG. 7, 21 is a zoom lens, 22 is an artificial retina,
Reference numeral 3 denotes a CCD image sensor, which is housed in one housing to form an artificial eyeball 24. Artificial eye 24
Is movably provided so that the field of view can be changed with respect to the object. The object image is formed on the entrance surface of the artificial retina 22 by the zoom lens 21, and the CCD image sensor 23
Receives an object image via the artificial retina 22 and converts the pattern into an electric signal. The CCD image sensor 23 is
As shown in FIG. 8, it has approximately 770 pixels in the horizontal direction and 490 pixels in the vertical direction, and the entire screen is divided into 63 areas of 9 areas horizontally and 7 areas vertically, and a portion corresponding to the central area. The artificial retinal cells are the central visual field cells 6 shown in FIG. 1, and the other artificial retinal cells are the peripheral visual field cells 7 shown in FIG. The output signal from the CCD image sensor 23 is converted into a digital signal by the A / D converter 25 and then taken into the computer 26. Of the captured image information, the information of the central portion is input to the neural network 27 for pattern recognition, and other information is used for peripheral information detection.

As shown in FIG. 9, the neural network 27 includes an input layer 80, an intermediate layer 90, and an output layer 10.
It has 0. CC corresponding to the central visual field cell 6 described above
The 85 × 70 pieces of image information of the central area of the D image sensor 23 are supplied to the input units 81, 82, ... Of the neural network 27, respectively. These pieces of information are output to the output unit 10 via the intermediate units 91, 92, ....
A signal representing the pattern of the object is output to 1, 102, .... Since the neural network 27 presents the object to be recognized in the central visual field 1a and is trained in advance according to the back propagation learning (back propagation) rule, the shape of the object can be discriminated from the output signal from the neural network 27. The signal representing the shape of the object is obtained in the next step
And further necessary information processing is performed.

If there is no object image in the central area, or
After recognizing the object in the central area, when it is necessary to determine another object, the peripheral information is used. Regarding the peripheral information, the integration / determination means 29 integrates the brightness for each area, compares the integrated value with a predetermined threshold value, and determines that an object exists in the area exceeding this threshold value. This determination information is supplied to the object selecting means 30. The object selecting means 30 randomly selects one area from the areas determined to have an object based on the random number generated by the random number generating means 31. The signal representing the selected region is the artificial eyeball driving means 32.
The artificial eyeball 24 is turned so that the selected object is in the center of the visual field.

As described above, the objects to be recognized can be selected one after another from the image including the plurality of objects to be recognized, and the objects can be distinguished and recognized.

If the neural network 27 cannot recognize the image, the magnification of the zoom lens 21 is changed to change the size of the image captured in the visual field so that the image is recognized again. It can also handle cases where the object does not fit, is too small, or when other objects are obstructing the recognition.

Further, it is clear that a two-dimensional map on an image can be drawn by combining the information of the neural network 27 and the position information of the object by the peripheral visual field cells. Furthermore, if this device is used two times. It can also be applied to the recognition of stereoscopic images. Of course, the neural network 27 does not have to be of the layered type described in this embodiment and of the type that uses the back propagation learning rule, but can be used for associative memory such as Hopfield model and associatron. All models can be used, and as shown in the figure, if the information of the neural network 27 is further transmitted to the subsequent processing,
It goes without saying that more advanced processing can be performed.

As described above, in the artificial retina of the present invention, since the artificial retinal cells, which are the constituent elements, are composed of the optical elements, the object information can be recognized in parallel and at high speed, and the separation and position detection processing can be performed. Further, since a lens having a substantially flat end face is used, it is possible to align each component with the lens end face as a reference, and it is not necessary to align them in the optical axis direction. In addition, the method for creating the 0th-order light cut part of the filter, which directly affects the detection function of the light-dark boundary and has the most severe alignment accuracy and which is a bottleneck for simplifying the device, by using the actual light flux is taken. Therefore, this accuracy can be greatly improved, and the device can be easily realized.

Furthermore, the artificial vision device of the present invention further comprises:
As described above, the artificial retina is divided into two functional parts, and the object to be recognized is made to be almost one in the central visual field part, so that the neural network for the recognition process is very easily constructed. be able to. Also, the next part to be recognized is judged from the information of the peripheral visual field part,
Since the central field of view is adjusted to that portion by the artificial eye movement mechanism, it becomes possible to recognize a plurality of objects in the image with the neural network having a simple configuration.

Although the artificial retina and the artificial vision device of the present invention have been described above based on the embodiments, the present invention is not limited to these embodiments and various modifications can be made.

[0035]

As is apparent from the above description, according to the artificial retina and the artificial vision device of the present invention, the artificial retina is
It consists of the first type artificial retinal cells that can recognize one object information and the second type artificial retinal cells that can separate multiple objects and detect their positions. Moreover, it is possible to select a specific item at high speed and easily recognize it.

[Brief description of drawings]

FIG. 1 is a diagram showing a structure of a central retinal cell constituting an artificial retina of the present invention.

FIG. 2 is a diagram showing a schematic configuration of an artificial retina according to the present invention.

FIG. 3 is a diagram showing the structure of peripheral retinal cells that form the artificial retina of the present invention.

FIG. 4 is a diagram for explaining the schematic configuration and operation of the artificial vision device according to the present invention.

FIG. 5 is a diagram for explaining movement of the visual field of the artificial vision device.

FIG. 6 is a diagram showing an example of an input image and an output image of the artificial retina.

FIG. 7 is a diagram showing a configuration of an artificial vision device according to a second embodiment.

FIG. 8 is a diagram for explaining a pixel array of a CCD image sensor and an example of screen division.

FIG. 9 is a diagram for explaining the configuration and operation of the neural network.

FIG. 10 is a diagram for explaining a human initial visual mechanism.

FIG. 11 is a diagram showing the excitation intensity of each human visual cell.

FIG. 12 is a circuit diagram of an example of a conventional initial vision mechanism.

FIG. 13 is a diagram for explaining a conventional method of optically detecting a bright / dark boundary.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Artificial retina 1a ... Central visual field 1b ... Peripheral visual field 2 ... Imaging means 3 ... Artificial eyeball 4 ... Neural network 5 ... Artificial eyeball moving means 6 ... Central visual field cell 7 ... Peripheral visual field cell 8 ... Detection means 9 ... Input surface 10 ... Fourier transform lens 10 11 ... Filter 11a ... 0th-order light cut portion 11b ... Higher-order light cut portion 11c ... Higher-order light cut portion 12 ... Output surface 13 ... Inverse Fourier transform lens 21 ... Zoom lens 22 ... Artificial retina 23 ... CCD image sensor 24 ... Artificial eyeball 25 ... A / D converter 26 ... Computer 27 ... Neural network 28 ... Next step 80 ... Input layer 81, 82, ... Input unit 90 ... Intermediate layer 91, 92, ... Intermediate unit 100 ... Output Layers 101, 102, ... Output unit 29 ... Integration / determination means 30 ... Object selection means 3 1 ... Random number generating means 32 ... Artificial eyeball driving means

Claims (3)

[Claims] 1. A Fourier transform lens for forming a Fourier transform image of an input object image on a predetermined surface, a filter arranged on the predetermined surface for removing a zero-order light component of the Fourier transform image, and filtering. Of the first kind of artificial retinal cell, which comprises an inverse Fourier transform lens for performing an inverse Fourier transform on the obtained Fourier transform image to return it to real space information, and a Fourier transform image for forming a Fourier transform image of the input object image on a predetermined surface. A second lens comprising a transform lens, a filter arranged on the predetermined surface for removing a higher-order light component of the Fourier transform image, and an inverse Fourier transform lens for performing an inverse Fourier transform on the filtered Fourier transform image to return it to real space information. An artificial retina comprising: a seed artificial retinal cell; and a detection element that detects output light from the first and second artificial retinal cells. 2. The Fourier transform lens for artificial retina cells of the first and second types has an entrance surface and an exit surface that are substantially flat, and the Fourier transform image is formed on the exit surface, and each of the filters Is placed in close contact with the exit surface, and the inverse Fourier transform lens of the first and second types of artificial retinal cells has an entrance surface and an exit surface that are substantially flat, and the entrance surface is the filter. The artificial retina according to claim 1, wherein the artificial retina is arranged so as to be in close contact with each other and is arranged so that an inverse Fourier transform image is formed on its exit surface. 3. An image forming means for forming an object image, an artificial eyeball comprising the artificial retina according to claim 1, and information received by the artificial retinal cells of the first type of the artificial retina. A neural network for pattern recognition of an object, a means for determining an object to be recognized next based on information detected by the second type artificial retinal cells of the artificial retina, and the artificial eyeball toward the recognition target. An artificial vision device comprising: an artificial eye movement device for moving the eye.