Pattern recognition optical system

Abstract

To provide a pattern recognition optical system capable of discriminating respective patterns simultaneously in parallel, even when a handwritten pattern or a pattern of different font exists at a discretionary position in the input image, and further when a plurality of these patterns are included.SOLUTION: A pattern recognition optical system 1 comprises: input image generation means 2 for generating an input image expressed by the intensity level or phase difference of light; an optical linear discriminant analysis filter 3 for performing a linear discriminant analysis using light by modulating the light from the input image generation means 2 and generating a plurality of focal spots at positions individually set for each of a plurality of classes to be discriminated, as well as causing a focal spot corresponding to a pattern included in the input image to emit brighter light than the other focal spots; and output image generation means 5 for generating an output image that is drawn by the focal spots generated by the optical linear discriminant analysis filter 3 and that includes a plurality of dots corresponding to each class to be discriminated.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to pattern recognition optical systems. [Background technology]

[0002] Pattern recognition is performed to recognize characters, objects, etc. contained in an image. Methods such as deep learning are widely used in pattern recognition, but the computation load on the computer is high, and power consumption and heat generation are problems. As a means of solving such problems, pattern recognition optical systems that use light waves to perform pattern recognition have attracted attention. For example, Patent Document 1 discloses a pattern recognition optical system that transmits transmitted light from an input image through a multi-layer filter and performs signal processing using light that mimics a neural network, thereby displaying the pattern discrimination results in an output image. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] U.S. Pat. No. 1,139,2830 Summary of the Invention [Problem to be solved by the invention]

[0004] The method used by the pattern recognition optical system of Patent Document 1 is D 2 This is called a Diffractive Deep Neural Network (NN), and unlike optical correlation calculation systems that use filters that react to specific patterns, it can distinguish patterns even if they are handwritten or the font changes. 2 Since NN performs signal processing that mimics a neural network using a large number of filters, in order to correctly recognize a pattern, there must be only one pattern present in the center of the input image; therefore, if the input image contains multiple patterns, the pattern cannot be recognized.

[0005] The present invention has been made against this background, and has as its object to provide a pattern recognition optical system that can simultaneously distinguish each pattern in parallel, even if a handwritten pattern or a pattern of different fonts is present at any position in an input image, or even if multiple such patterns are included. [Means for solving the problem]

[0006] In order to achieve the above object, a pattern recognition optical system according to the present invention comprises: A pattern recognition optical system for determining to which of a plurality of classes a pattern included in an input image belongs, comprising: an input image generating means for generating the input image expressed by light intensity or phase difference; an optical linear discriminant analysis filter that performs linear discriminant analysis using light by modulating the light from the input image generating means, generates a plurality of focused spots at positions individually set for a plurality of classes to be discriminated, and causes a focused spot corresponding to a pattern included in the input image to emit light brighter than other focused spots; an output image generating means for generating an output image including a plurality of dots corresponding to each class to be discriminated, the output image being drawn by the focused light spots generated by the optical linear discriminant analysis filter; Equipped with.

[0007] The optical system may further include a lens disposed between the optical linear discriminant analysis filter and the output image generating means, the lens focusing the light from the optical linear discriminant analysis filter and outputting the light toward the output image generating means.

[0008] the input image generating means, the optical linear discriminant analysis filter, the lens, and the output image generating means are arranged in sequence at equal intervals; a front focal length of the optical linear discriminant analysis filter and a focal length of the lens are approximately equal to the interval; The back focal length of the optical linear discriminant analysis filter may be infinity.

[0009] In the optical linear discriminant analysis filter, a plurality of photon sieves each having a plurality of pinholes for transmitting light therethrough may be superimposed on a light-shielding film.

[0010] The input image generating means may transmit or reflect light from a light source in accordance with a pattern contained in the input image.

[0011] The output image generating means may be an imaging means for capturing an output image including dots drawn by focused light spots generated by the optical linear discriminant analysis filter. Effect of the Invention

[0012] According to the present invention, it is possible to provide a pattern recognition optical system that can simultaneously and in parallel distinguish between handwritten patterns or patterns of different fonts present at any position in an input image, and even when multiple such patterns are included. [Brief description of the drawings]

[0013] [Figure 1] 1A and 1B are diagrams showing examples of an input image and an output image according to the first embodiment of the present invention. [Diagram 2] 1 is a diagram showing a configuration of a pattern recognition optical system according to a first embodiment of the present invention. [Diagram 3] FIG. 2 is a diagram showing an example of a transmissive and reflective input image generating means according to the first embodiment of the present invention. [Figure 4] FIG. 2 is a diagram showing a state in which an input image according to the first embodiment of the present invention is divided into a plurality of pixels. [Diagram 5] FIG. 2 is a diagram illustrating a mechanism of an optical linear discriminant analysis filter according to the first embodiment of the present invention. [Figure 6] FIG. 2 is a diagram showing a mechanism of an input image filter according to the first embodiment of the present invention. [Figure 7] FIG. 1 is a diagram showing the mechanism of a photon sieve according to a first embodiment of the present invention. [Figure 8]4 is a flowchart showing a flow of an OLDF design process according to the first embodiment of the present invention. [Figure 9] 5 is a flowchart showing the flow of observation vector generation processing according to the first embodiment of the present invention. [Figure 10] 5 is a flowchart showing a flow of input image filter design processing according to the first embodiment of the present invention. [Figure 11] FIG. 2 is a diagram showing a state in which two photon sieves according to the first embodiment of the present invention are combined. [Figure 12] 4 is a flowchart showing the flow of a transmittance optimization process according to the first embodiment of the present invention. [Figure 13] FIG. 13 is a diagram showing the relationship between a shift of an input image and a shift of a focused spot. [Figure 14] FIG. 11 is a diagram showing a configuration of a pattern recognition optical system according to a second embodiment of the present invention. [Figure 15] FIG. 1(a) is a diagram showing how a photon sieve according to a second embodiment of the present invention generates a plane wave, and FIG. 1(b) is a diagram showing how the optical path of the photon sieve in (a) is reversed. [Figure 16] FIG. 1A is a diagram showing how a photon sieve according to a second embodiment of the present invention focuses a plane wave, and FIG. 1B is a diagram showing how the focusing point in FIG. [Figure 17] FIG. 10(a) is a diagram showing how a lens according to a second embodiment of the present invention focuses a plane wave, and (b) is a diagram showing how the optical path of the lens in (a) is reversed. [Figure 18] FIG. 1(a) is a diagram showing the optical path through an optical system having a photon sieve and a lens in accordance with a second embodiment of the present invention, and FIG. 1(b) is a diagram showing the optical path when two photon sieves with different pinhole shift amounts are overlapped in the optical system of (a). DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, an embodiment of a pattern recognition optical system according to the present invention will be described in detail with reference to the drawings. In each drawing, the same or equivalent parts are denoted by the same reference numerals.

[0015] (Embodiment 1) First, a description will be given of a pattern recognition optical system according to embodiment 1. The pattern recognition optical system according to embodiment 1 is a system for determining a pattern displayed in an input image based on the positions of dots displayed in an output image.

[0016] The input image is, for example, a grayscale or binarized image, and includes one or more patterns. The patterns include, for example, characters, numbers, symbols, and figures. Patterns of the same type may be in different fonts or may be handwritten. The characters may be in any language, for example, hiragana, katakana, kanji, or the alphabet. The figures may be drawn on a plane or may be structures including projections and recesses. The patterns included in the input image are drawn within a frame set at a specific position of the input image. As an example, FIG. 1 illustrates a specific example in which numbers are discriminated as patterns to be discriminated, and two types of patterns, the numbers 3 and 7, are drawn in the lower left and upper right of the input image.

[0017] In the output image, a dot group consisting of multiple dots that correspond to the pattern in the input image is drawn, and the brightest dot in the dot group corresponds to the pattern in the input image. The dot group is generated at a position that is vertically and horizontally inverted with respect to the position of the pattern in the input image. The vertical and horizontal inversion is due to the use of a photon sieve in the optical system. The number of dots in each dot group is the same as the number of patterns to be identified. The arrangement of the dots in the dot group can be arbitrary.

[0018] For example, in the upper right and lower left frames of the output image shown in Fig. 1, a group of dots corresponding to a pattern of 3 and a group of dots corresponding to a pattern of 7 are drawn, respectively. In each group of dots, 10 dots are arranged corresponding to the 10 numbers from 0 to 9. In the group of dots corresponding to the pattern of 3, the rightmost dot in the bottom row of the 10 dots is the brightest, and in the group of dots corresponding to the pattern of 7, the leftmost dot in the third row from the bottom is the brightest.

[0019] 2, the pattern recognition optical system 1 includes an input image generating means 2, an optical linear discriminant analysis filter 3 that generates a focused spot by diffracting and interfering light indicating the pattern of the input image emitted by the input image generating means 2, a lens 4 that focuses the light emitted by the optical linear discriminant analysis filter 3, and an output image generating means 5 that generates an output image in which dots are drawn by the light from the lens 4. The components of the pattern recognition optical system 1 are arranged in the order of the input image generating means 2, the optical linear discriminant analysis filter 3, the lens 4, and the output image generating means 5.

[0020] The input image generating means 2 is a means for generating an input image showing a pattern by the intensity of light. The input image generating means 2 includes a laser light source and a generating means for generating an input image showing a pattern by transmitting or reflecting light from the laser light source. The generating means is formed with transparent parts that transmit light and non-transparent parts that do not transmit light according to the shape of the pattern.

[0021] As shown in Fig. 3, the generating means may be, for example, a transmission type in which a transmission region is formed according to the shape of the pattern, or a reflection type in which a reflection region is formed according to the shape of the pattern. In the case of a reflection type generating means, a convex portion may be provided according to the shape of the pattern to change the phase of light between the pattern portion and the other portion. The wavelength of the light from the laser light source is not particularly limited, and any electromagnetic wave in any wavelength range may be used as long as it is a coherent electromagnetic wave.

[0022] As shown in Figure 4, the input image is divided into multiple pixels for convenience, and each pixel is treated as a point light source with light intensity or phase difference. The input image is divided into vertical and horizontal grids, and each section is a pixel. The brightness of a point light source can be quantified by the amplitude of the light emitted by the point light source. For example, the complex amplitude x of the light emitted from the nth pixel is n When a transmissive filter is used, the amplitude of light from a point light source becomes relatively large in pixels that correspond to the pattern, and the amplitude of light from the point light source becomes small in pixels that do not correspond to the pattern.

[0023] The optical linear discriminant analysis filter 3 is abbreviated as OLDF (Optical Linear Discriminant Filter), and is a filter that generates a plurality of focused spots on the same plane on the opposite side of the input image as shown in Fig. 5 when light from a point light source distributed in accordance with the pattern of the input image passes through it. The optical linear discriminant analysis filter 3 is composed of a light-shielding film having a large number of pinholes provided therein. The light-shielding film is, for example, a thin metal film.

[0024] The number of focused spots generated by the optical linear discriminant analysis filter 3 is equal to the number of types of patterns to be discriminated, and each focused spot is associated with a pattern to be discriminated. In more detail, light is focused so that the focused spot corresponding to the pattern included in the input image among the multiple generated focused spots is the brightest.

[0025] The optical linear discriminant analysis filter 3 is a combination of all input image filters corresponding to each pattern to be discriminated, for each frame in which a pattern set in the input image is drawn. In the optical linear discriminant analysis filter 3, when combining the input image filters, the positions of the pinholes are shifted so that the focused spots corresponding to each pattern are generated in different locations.

[0026] As shown in Figure 6, the input image filter is composed of a light-shielding film with many pinholes, and when light from point light sources (pixels) distributed according to the pattern of the input image passes through it, it generates a single focused spot at a specific position on the output image. The input image filter is composed of a photon sieve that acts like a lens by itself through the diffraction and interference of light. By arranging the pinholes so that they do not overlap, multiple photon sieves can be stacked on one light-shielding film, making it possible to combine multiple input image filters into a single filter.

[0027] As shown in FIG. 7, the photon sieve is an optical element in which multiple pinholes are provided on one or multiple ellipses set on a light-shielding film. The shape of the ellipse is obtained by determining the locus of point R that satisfies the condition that the sum of the distance from the point light source to point R on the filter and the distance from point R to the focused spot is constant. Multiple ellipses can be defined by allowing a deviation of an integer multiple of the wavelength of the illumination light from the sum of the above distances. The phase difference φ between the point light source and the focused spot can also be freely set by adjusting the sum of the above distances. The pinholes on the same ellipse are arranged at equal intervals. Since the photon sieve has the above configuration, it focuses light emitted from a point light source at a point that is a front focal length f1 away from the photon sieve, and generates a focused spot at a point that is a back focal length f2 away from the photon sieve.

[0028] The position of the focused spot and the focal lengths f1 and f2 can be set arbitrarily by changing the size and number of the pinholes. Also, the more ellipses the pinholes are machined into, the brighter and smaller the focused spot can be. If the amplitude of light at the point light source is x, the amplitude of light at the focused spot is A0wx. A0 is the amplitude of the focused spot when the amplitude of the point light source is 1 in the reference photon sieve. The reference photon sieve is a photon sieve with the smallest pinhole spacing that is structurally feasible, and details will be described later. w is the transmittance (proportionality constant) of the photon sieve, and can be controlled by the density of the pinholes.

[0029] Returning to Fig. 2, the lens 4 focuses the light emitted from the optical linear discriminant analysis filter 3 and emits it towards the output image generating means 5. It is preferable that the lens 4 is free of aberration.

[0030] The output image generating means 5 is a means for projecting light from the lens 4 and generating an output image composed of a plurality of dots. The output image generating means 5 is, for example, a screen or a camera (imaging means). In the output image, a plurality of dots are drawn by focused spots generated by a plurality of input image filters synthesized in the optical linear discriminant analysis filter 3. The user may refer to the output image projected on the screen and determine that a pattern corresponding to the brightest dot in each dot group is displayed in the input image. In addition, a computer may measure the brightness of each dot in an image acquired by the camera, and when the measured value is equal to or greater than a preset threshold value, determine that a pattern corresponding to a measured value equal to or greater than the threshold value is generated in the input image. The above is the configuration of the pattern recognition optical system 1.

[0031] (OLDF design processing) Next, the OLDF design process according to the first embodiment will be described with reference to the flowchart in Fig. 8. The OLDF design process is a process for designing the optical linear discriminant analysis filter 3 by synthesizing a plurality of photon sieves that diffract and interfere with light showing a pattern of an input image to generate a focused spot.

[0032] First, an observation vector generation process is executed to generate a large number of observation vectors for discriminating patterns included in an input image based on learning data of images including the patterns (step S1). Hereinafter, the flow of the observation vector generation process according to the first embodiment will be described with reference to the flowchart in FIG.

[0033] (Observation vector generation process) First, a plurality of image data are prepared as learning data for each pattern to be discriminated (step S11). The image data includes one pattern per image, and the pattern is arranged in the center of the image. It is preferable that the input image has as many variations as possible for one type of pattern. For example, when numbers or letters are used as patterns, it is good to prepare patterns handwritten by multiple people and in different fonts.

[0034] Next, each image is divided into a plurality of pixels (step S12). Specifically, first, the image is divided into a plurality of pixels using a vertical and horizontal grid as shown in Fig. 4. Here, the number of pixels in the input image is set to N.

[0035] Next, a vector X representing an input pattern of the image is generated based on the image divided into a plurality of pixels (step S13). Specifically, first, the complex amplitude x n (1≦n≦N). Next, find the N complex amplitudes x n By arranging these, a vector X that represents the pattern of the input image is generated. X = [x1,x2,…,x N ] T …(1)

[0036] The bright parts of the input image have complex amplitude x n becomes larger, and the dark areas are the complex amplitude x n By reducing |x, the input image can be quantitatively expressed by the vector X that indicates the pattern of the input image. n The whole is normalized so that the largest | is 1. Below, vector X may be referred to as the "input pattern."

[0037] Next, a linear discriminant analysis is performed based on a large number of input patterns X generated for each type of pattern, and an observation vector W corresponding to each pattern to be discriminated is generated. (C) (step S14), and the process returns. (C)is a vector of complex numbers used to determine whether an input pattern X belongs to a particular class C, and is expressed as follows. Its dimension is the same as the number of pixels N in the input image. W (C) = [w1, w2, …, w N ] T (|w n |<1) …(2)

[0038] Class C corresponds to the type of pattern in the input image. (C) The number of observation vectors W is required to correspond to the type of pattern to be discriminated. For example, if the pattern to be discriminated is a number from 0 to 9, 10 observation vectors W are required. (C) Prepare the following.

[0039] Observation vector W (C) is calculated using linear discriminant analysis. Specifically, vectors X that indicate the same type of pattern are X1 to X2 even if they are patterns written by different people or in different fonts. N In a space with coordinate axes, the vectors X and W are distributed close to each other. On the other hand, input patterns X that show different types of patterns are distributed in different places. For this reason, multiple vectors X that show different types of patterns can be distinguished from each other by setting an N-1-dimensional hyperplane (decision boundary). This decision boundary is defined by the observation vector W (C) The observation vector W (C) is a vector of length 1 perpendicular to the decision boundary. To determine the class C of an unknown image, first, we first define the observation vector W (C) and the input pattern X, and then calculate the inner product W (C)T X and the distance w0 from the origin to the decision boundary are compared with each other.

[0040] As the linear discriminant analysis, for example, Fisher's linear discriminant analysis may be used. Fisher's linear discriminant analysis is a method for determining the observation vector W so as to maximize the ratio of the between-class variance to the within-class variance. (C) This is a method for determining the The above is the flow of the observation vector generation process.

[0041] The observation vector W generated in the above process (C) When an optical linear discriminant analysis filter 3 is generated using (C) The brightness of the dots drawn on the output image is determined by the observation vector W (C) and the input pattern X, the absolute value of which is squared |W (C)T X| 2 If the brightness of the dots in the output image is brighter than a preset threshold, the input pattern X can be determined to belong to class C. Note that the observed vector W (C) The inner product of and input pattern X is expressed by the following equation. W (C)T X=w * 1x1+w * 2x2+…+w * N x N …(3)

[0042] Returning to FIG. 8, each observation vector W generated in step S1 (C) Then, an input image filter design process is executed to generate a plurality of input image filters corresponding to the input image filters (step S2). Hereinafter, the flow of the input image filter design process according to the first embodiment will be described with reference to the flowchart in FIG.

[0043] (Input image filter design process) First, a point light source s n Photon sieve P with different coordinates (n=1,2,…,N) and the same focal spot coordinate Q n The coordinates of the pinholes (n=1, 2, ..., N) are set (step S21). A specific example is shown in Fig. 11. In Fig. 11, for ease of understanding, two photon sieves P1 and P2 corresponding to two point light sources s1 and s2 are shown.

[0044] In the process of step S21, the coordinates of the point light sources s1, s2, ... and the coordinate Q of the focused spot are aligned with each photon sieve P nThe coordinates of the pinhole in the point light source s are determined. n The optical path length from the coordinates (x0, y0, z0) of the target focused spot through the pinhole to the coordinate Q (x1, y1, z1) of the target focused spot is set to be constant regardless of which pinhole is passed through. The ellipses on which the pinholes are placed may be different in size.

[0045] The photon sieve to be designed is a reference photon sieve with the minimum pinhole interval that is structurally feasible. The reference photon sieve is (C) There is only one, regardless of the number of or the number of pixels N. Note that the amplitude of the point light source in the reference photon sieve is 1, and the amplitude of the focused spot is A0.

[0046] Next, each photon sieve P n Transmittance t n (Step S22) The transmittance optimization process according to the first embodiment will be described below with reference to the flowchart in FIG.

[0047] First, the transmittance t n The initial value of t n =w * n / |w * n (step S31). * n is the observation vector W generated in step S1. (C) is a component of.

[0048] Next, each photon sieve P n The corresponding point light source s n (n=1,2,…,N), the amplitudes of each are calculated based on the input pattern X generated in step S1 as x1, x2,…, x N (Step S32). n The light from each of the photon sieves P m(m=1,2,...,N) to generate a total of N focused spots. These focused spots have a certain degree of spread. In the specific example of Figure 11, the light from point sources s1 and s2 also passes through two photon sieves P1 and P2, respectively, to generate two focused spots each.

[0049] point light source n All the light from the Photon Sieve P m The amplitude of the focused spot at the coordinate Q when passing through is a nm Then, the following simultaneous equations are solved (step S33).

[0050]

number

[0051] Amplitude a nm can be calculated using the following formula:

[0052]

number

[0053] In the above formula, M is the photon sieve P m The total number of pinholes in (x p ,y p ) is Photon Sieve P m represents the coordinates of the pinhole contained in

[0054] Next, in step S33, the transmittance t n It is determined whether or not the transmittance t' that was set before solving the simultaneous equations is converged (step S34). n For |t n -t' n | / |t n When | is smaller than a predefined tolerance ε, it is determined that the solution has converged. The smaller the tolerance ε is set to, the higher the accuracy of the pattern recognition process. nIf it is determined that the transmittance t n If it is determined that the solution of the simultaneous equations, the transmittance t n Photon Sieve P n Transmittance t n (step S35), and the process returns to step S33. The above is the flow of the transmittance optimization process.

[0055] Returning to FIG. 10, the transmittance t optimized in the transmittance optimization process (step S22) in FIG. n Based on each photon sieve P n The pinhole spacing is adjusted (step S23). Specifically, the pinhole spacing is adjusted while maintaining the condition of equal spacing, and the pinholes are thinned out so that the total number of pinholes is A / A0 times the total number of pinholes in the reference photon sieve. Then, the amplitude of the focused spot changes from A0 to A0A (A<1). The coefficient A and the transmittance t n and satisfy the following formula: φ is the phase difference between the point light source and the focused spot described above. t n =Aexp(-jφ) …(6) From the above formula, the transmittance t n It is understood that the above can be adjusted.

[0056] Next, it is determined whether the pinholes overlap (step S24). If it is determined that the pinholes overlap (step S24; Yes), the pinholes existing on the same ellipse are moved on the ellipse while maintaining the intervals between them, and the pinhole overlap is eliminated as much as possible (step S25), and the process returns. On the other hand, if it is determined that the pinholes do not overlap (step S24; No), the process ends as it is. This allows the coordinate data of the pinholes of the input image filter that generates the focused spots according to the pattern as shown in FIG. 6 to be obtained. The above is the flow of the input image filter design process.

[0057] After the above processing, the amplitude a at the coordinate Q of the focused spot (observation point) is expressed by the following formula. a=A0w * 1x1+A0w * 2x2+…A0w * N x N =A0W T X …(7) Therefore, the brightness of the focused spot at the coordinate Q of the focused spot is |A0| 2 |W T X| 2 However, the proportionality constant |A0| 2 does not affect the accuracy of pattern recognition. The squared absolute value of the amplitude a at a certain focused spot, |a| 2 is the threshold |w0| 2 If this is the case, it can be determined that the pattern corresponding to the focused spot is included in the input screen.

[0058] Returning to FIG. 8, each observation vector W generated in step S2 (C) The optical linear discriminant analysis filter 3 is generated by synthesizing a plurality of input image filters corresponding to the above (step S3). At this time, the coordinate Q of the focused spot is calculated based on the observation vector W (C) The synthesis is done so that each one is different.

[0059] Next, it is determined whether the pinholes overlap on the same plane (step S4). If it is determined that the pinholes overlap (step S4; Yes), the pinholes existing on the same ellipse are moved on the ellipse while maintaining the intervals, and the pinhole overlap is eliminated as much as possible (step S5), and the process ends. On the other hand, if it is determined that the pinholes do not overlap (step S4; No), the process ends as it is. As a result, the coordinate data of the pinholes that constitute the optical linear discriminant analysis filter 3 that generates focused spots according to the number of patterns as shown in FIG. 5 is obtained. The above is the flow of the OLDF design process.

[0060] Then, a pinhole is processed on one light-shielding plate based on the pinhole coordinate data generated by the OLDF design process. The pinhole can be processed by, for example, a technique such as etching. This creates an optical linear discriminant analysis filter that generates focused spots according to the number of patterns to be discriminated.

[0061] As described above, the pattern recognition optical system 1 according to the first embodiment includes an input image generating means 2 that generates an input image represented by the intensity of light, an optical linear discriminant analysis filter 3 that performs linear discriminant analysis by modulating the light from the input image generating means 2, and an output image generating means 5 that generates an output image including a plurality of dots corresponding to each class using the light from the optical linear discriminant analysis filter 3.

[0062] Therefore, even if an input image contains multiple patterns, each pattern can be distinguished. In addition, because patterns can be distinguished simply by operating the laser light source, power consumption and heat generation can be reduced, and processing time and power consumption remain unchanged regardless of the number of patterns to be distinguished in the input image.

[0063] (Embodiment 2) Next, a pattern recognition optical system 1 according to embodiment 2 will be described. In the pattern recognition optical system 1 according to embodiment 1, the brightness of the focused spot changes when the position of the input image shifts, whereas in the pattern recognition optical system 1 according to embodiment 2, the brightness of the focused spot does not change even when the position of the input image shifts. The following description will focus on the differences between the two systems.

[0064] In the pattern recognition optical system 1 according to the first embodiment, when the position of the input image is shifted as shown in FIG. 13, the inner product W (C)T X cannot be calculated correctly, and the phases of the focused spots created by each photon sieve change randomly. This causes the brightness of the focused spots to change. In the pattern recognition optical system 1 according to the second embodiment, even if the point light source shifts, the amplitude of each spot in the output image does not change, and |WT X| 2 The configuration is such that the condition that the value of does not change (isoplanatic condition) is satisfied.

[0065] In the pattern recognition optical system 1 according to the second embodiment, the front focal length of the optical linear discriminant analysis filter 3 is z0, the rear focal length is ∞, and the focal length of the lens 4 is z0, and the input image generating means 2, the optical linear discriminant analysis filter 3, the lens 4, and the output image generating means 5 are each disposed at an equal interval z0, as shown in FIG. 14.

[0066] The optical linear discriminant analysis filter 3 is made by combining multiple photon sieves with a front focal length of z0 and a back focal length of infinity. As shown in Figure 15(a), in a photon sieve with a front focal length of z0 and a back focal length of infinity, when light from a point light source passes through the photon sieve, it becomes a plane wave. To design such a photon sieve, it is useful to consider the direction of the light rays in reverse. More specifically, as shown in Figure 15(b), x ,k y ) is incident on a photon sieve and focused on a point (x0, y0, z0). If the light path is reversed in this photon sieve, a spherical wave from a point source at (x0, y0, z0) passes through the photon sieve with a wave vector (-k x ,-k y ) plane wave.

[0067] In such a photon sieve, if the phase of the incident wave at the origin is 0, the phase at the focal point is -jk x x-jk y The equation is y-jkr (j is the imaginary unit of a complex number). The locus of (x, y, 0) where this is constant (however, deviations of an integer multiple of 2π are allowed) is found, and a pinhole is placed on that locus. Then, all the light waves that pass through the pinhole and reach the focal point (x0, y0, z0) reinforce each other in the same phase, and a bright focal spot appears.

[0068] In the pattern recognition optical system 1, the amplitude a of the focused spot before movement is expressed by the following formula. a=w1x1+w2x2+…w N x N …(8) On the other hand, the amplitude a' of the focused spot after the movement is expressed by the following formula. L is the movement distance of the spot. Even if the input image is shifted, the phase change amount e jθ is constant regardless of the composition. a'=e jθ w1x1+e jθ w2x2+…+e jθ w N x N =e jθ [w1x1+w2x2+…+w N x N ] …(9) where θ=-πL 2 / (λz0).

[0069] From the above equation, the brightness of the focused spot before and after movement has the following relationship, and it can be understood that the amplitude (brightness) of the dot remains unchanged even if the focused spot is shifted. |a| 2 =|a'| 2 …(10) The above is the configuration of the pattern recognition optical system 1.

[0070] In the following, with reference to Figs. 16 to 18, we verify whether the photon sieve designed by the above procedure is isoplanatic. First, consider a photon sieve with an incident wave wave vector (0,0) and a focused spot (0,0,z0) as shown in Fig. 16(a). In this photon sieve, the wave vector of the input wave is (Δk x , 0) and shifting the pinhole by Δx, the position of the focused spot shifts to (Δx, 0, 0) as shown in Figure 16(b), and the phase changes by Δφ. Δφ is expressed by the following equation. Δφ=-k / (2z0)(Δx0 2 -Δx 2 ) …(11)

[0071] Considering the light path in reverse, if the pinhole is shifted by Δx and the point light source is shifted from x=0 to Δx0, the wave number of the plane wave obtained after passing through the photon sieve is Δk x and the phase of the plane wave at the origin changes by Δφ.

[0072] Now consider Lens 4, where aberration has been removed. In Lens 4, light rays incident parallel to the optical axis are focused at the origin, as shown in Figure 17(a). If we consider the optical path in reverse, in Lens 4, light rays that pass through the origin are focused at a focusing point on the focal plane, as shown in Figure 17(b). And no matter in which direction the light rays propagate from the origin, the phase of the focusing point on the focal plane does not change. In other words, the phase of the focusing point is determined only by the phase when it passes through the origin.

[0073] For the above reasons, if the phase of the plane wave transmitted through the photon sieve shifts by Δφ at the origin as shown in Figure 18(a), the spot phase on the focal plane will also shift by the same Δφ. Now, let's try overlapping two photon sieves P1 and P2 as shown in Figure 18(b). The photon sieve P1 has a pinhole shift amount of Δx1, and the photon sieve P2 has a pinhole shift amount of Δx2. In this case, the phases φ1 and φ2 before the point light source shift are as follows: φ1=-k / (2z0)Δx1 2 …(12) φ2=-k / (2z0)Δx2 2 …(13)

[0074] On the other hand, the phases φ1 and φ2 before the point light source shift are as follows: φ1=-k / (2z0)(Δx1 2 -Δx0 2 ) …(14) φ2=-k / (2z0)(Δx2 2 -Δx0 2 ) …(15)

[0075] According to equations (12) to (15), the phase difference φ2-φ1 remains constant even if the point light source shifts by Δx0. From the above, it can be seen that the pattern recognition optical system 1 according to the second embodiment satisfies the isoplanatic condition.

[0076] As described above, in the pattern recognition optical system 1 according to the second embodiment, the input image generating means 2, the optical linear discriminant analysis filter 3, the lens 4 and the output image generating means 5 are arranged in order at equal intervals, the front focal length of the optical linear discriminant analysis filter 3 and the focal length of the lens 4 approximately match the intervals, and the rear focal length of the optical linear discriminant analysis filter 3 is infinity.

[0077] This means that the pattern can be identified no matter where it is in the image, and for example, it can be recognized even if the pattern is not at the center of the input image. 2 Unlike NNs, there is no need for preprocessing such as adjusting the position of the target pattern so that it is at the center of the input image, or extracting the target patterns one by one from the image and inputting them into the system. Since the arrangement of multiple patterns in the input image is maintained in the output image, it is effective for processing patterns that have important meanings (for example, words made up of arranged characters).

[0078] The present invention is not limited to the above-described embodiment, and the following modifications are possible.

[0079] (Modification) In the above embodiment, the lens 4 is disposed between the optical linear discriminant analysis filter 3 and the output image generating means 5, but the present invention is not limited to this. For example, the lens 4 may be omitted depending on the size of the output image generating means 5.

[0080] In the above embodiment, an output image is generated from an input image including a plurality of patterns, and the plurality of patterns are determined based on the dots drawn in the output image, but the present invention is not limited to this. The input image may include only one pattern.

[0081] In the above embodiment, Fisher's linear discriminant analysis is used to calculate the observation vector W (C) However, the present invention is not limited to this. For example, the observation vector W (C) may be generated.

[0082] In the above embodiment, the optical linear discriminant analysis filter 3 is irradiated with laser light, but the present invention is not limited to this. Any electromagnetic wave having a coherent wavelength may be irradiated, for example, microwaves, terahertz waves, infrared rays, visible light, ultraviolet rays, or X-rays.

[0083] In the above embodiment, the pattern in the input image including the pattern is discriminated, but the present invention is not limited to this. The object may be directly illuminated and the reflected light may be input to the filter to discriminate the three-dimensional structure of the object. For example, the present invention may be used for product inspection on an assembly line, the vision of an autonomous robot, and the control of automatic driving.

[0084] In the above embodiment, the user executes the OLDF design process using a computer, but the present invention is not limited to this. For example, a computer having a memory and a processor may be prepared, and the above process may be executed by the processor executing a program stored in the memory. The present invention may also be realized as a program, or as a storage medium on which the program is recorded.

[0085] The program for executing the above-mentioned processing operations may be stored in a computer-readable recording medium such as a flexible disk, a CD-ROM (Compact Disk Read-Only Memory), a DVD (Digital Versatile Disk), an MO (Magneto-Optical Disk), etc. In this case, the program stored in the recording medium is installed in a computer to configure a processor that executes the above-mentioned processing.

[0086] The above-mentioned embodiments are merely examples, and the present invention is not limited to these, and various embodiments are possible without departing from the spirit of the invention described in the claims. The components described in each embodiment and modification can be freely combined. In addition, inventions equivalent to the inventions described in the claims are also included in the present invention. [Explanation of symbols]

[0087] 1. Pattern Recognition Optical System 2. Input image generation means 3. Optical Linear Discriminant Analysis Filter 4 Lenses 5. Output image generation means

Claims

1. A pattern recognition optical system for determining to which of a plurality of classes a pattern included in an input image belongs, comprising: an input image generating means for generating the input image expressed by light intensity or phase difference; an optical linear discriminant analysis filter that performs linear discriminant analysis using light by modulating the light from the input image generating means, generates a plurality of focused spots at positions individually set for a plurality of classes to be discriminated, and causes a focused spot corresponding to a pattern included in the input image to emit light brighter than other focused spots; an output image generating means for generating an output image including a plurality of dots corresponding to each class to be discriminated, the output image being drawn by the focused light spots generated by the optical linear discriminant analysis filter; A pattern recognition optical system comprising:

2. a lens disposed between the optical linear discriminant analysis filter and the output image generating means, the lens focusing the light from the optical linear discriminant analysis filter and outputting the light toward the output image generating means; 10. The pattern recognition optical system of claim 1.

3. the input image generating means, the optical linear discriminant analysis filter, the lens, and the output image generating means are arranged in sequence at equal intervals; a front focal length of the optical linear discriminant analysis filter and a focal length of the lens are approximately equal to the interval; The rear focal length of the optical linear discriminant analysis filter is infinity; 3. The pattern recognition optical system of claim 2.

4. In the optical linear discriminant analysis filter, a plurality of photon sieves each having a plurality of pinholes for transmitting light are superimposed on a light-shielding film.

4. A pattern recognition optical system according to any one of claims 1 to 3.

5. The input image generating means transmits or reflects light from a light source in accordance with a pattern included in the input image.

4. A pattern recognition optical system according to any one of claims 1 to 3.

6. the output image generating means is an imaging means for capturing an output image including dots drawn by the focused light spots generated by the optical linear discriminant analysis filter; 4. A pattern recognition optical system according to any one of claims 1 to 3.

Images