JPH0527300A - Camera having learning function - Google Patents

Abstract

PURPOSE:To provide a camera having a learning function for increasing operation speed and a correct answer ratio in a camera using a neural network. CONSTITUTION:In this camera having plural photometric sensors (a), an exposure control means (b) controlling the exposure of the camera based on the output signal of the photometric sensor (a), a network (c) capable of executing learning by a self-organization so that an exposure control signal is calculated from the output signal of the photometric sensor (a), and sent out to the exposure control means (b), and a learning function provided with a learning means (d) self-organizing the network (c) so that the output value of the network (c) is made incident with a desired output value, a converting means (e) is interposed between the photometric sensor (a) and the network (c), and plural output signals from the photometric sensor (a) are converted into parameters which are less than the output signals and inputted to the network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a camera having a learning function capable of learning a photographer's favorite exposure control method and performing exposure control based on the learning result.

[0002]

2. Description of the Related Art In a conventional camera, one of a plurality of modes is selected on the basis of a predetermined program or a photographer's instruction, and exposure control is performed by an exposure control method determined by this mode. However, there is a problem in that each mode gives an optimum exposure result only to a limited subject pattern, and an unintended result can be obtained if the mode is selected incorrectly. On the other hand, when trying to give desired exposure correction to all subject patterns without setting the mode, there are problems that formulation of parameters is difficult and the program becomes huge and the calculation time becomes long. Further, there is a problem that it is difficult to satisfy the tastes of all photographers depending on the uniform mode or program.

Therefore, the output values from the photosensors in matrix form are input to the neural network, the neural network is learned so that the output value desired by the photographer can be obtained, and the neural network in which self-organization by learning is completed. A camera with a learning function has been proposed in which the output value from the optical sensor is input to the camera and the output value from the network is used for exposure control (Japanese Patent Laid-Open No. 2-96).
724).

[0004]

However, in the camera with the learning function described above, since the plurality of outputs obtained from the matrix-shaped photosensors are directly input to the neural network, the number of variables to be input does not increase. In many cases, there was a problem that the network takes a long time to learn.

There are also unnecessary variables that do not directly contribute to the output of the network among the variables that are input, and this makes learning time extremely long and, of course, due to the nature of the neural network. However, if an unnecessary variable or a variable that requires an inconsistent answer is input during learning, the learning result may not converge or a correct value may not be output.

Further, if there are many input variables, the scale of the network is inevitably large, and there is a problem that the calculation time until the exposure is determined may become long in the actual photographing.

An object of the present invention is to provide, in a camera using a neural network, a camera having a learning function capable of speeding up calculation and improving a correct answer rate.

[0008]

With reference to FIG. 1, which is a diagram corresponding to claims, the present invention is based on a plurality of photometric sensors a for outputting photometric information and an output signal of the photometric sensors a. Exposure control means b for controlling the exposure of the camera,
In order to match the output value of the network c with the network c capable of learning by self-organization, which outputs the exposure control signal to the exposure control means b using the output signal of the photometric sensor a as an input parameter, to a desired output value. Further, it is applied to a camera having a learning function provided with a learning means d for self-organizing the network c. Further, the above-mentioned object is provided between the photometric sensor a and the network c, and provided with a conversion means e for converting a plurality of output signals from the photometric sensor a into a smaller number of network input parameters. It is achieved by

[0009]

The output signal from the photometric sensor a is converted by the conversion means e into a smaller number of network input parameters than the output signal, and is input to the network c. The network c self-organized by the learning means d outputs an exposure control signal based on this input parameter, and the exposure control means b controls the exposure of the camera based on the signal from this network c.

[0010]

【Example】

First Embodiment FIG. 2 is a diagram showing a first embodiment of a camera having a learning function according to the present invention. In FIG. 2, in FIG.
Is a photometric device, 2 is A / of the photometric output from the photometric device 1.
An A / D converter for performing D conversion, 3 is a neuro computer including a neural network, memory, etc., 4 is a computing device for computing an exposure value, 5 is an exposure control device, and a shutter (not shown) for controlling a shutter and a shutter. The mechanism unit 6 and the diaphragm mechanism unit 7 are provided respectively.

In the present embodiment, the subject screen is divided into five areas, that is, a central portion, an upper left portion, an upper right portion, a lower left portion, and an upper left portion,
As shown in FIG. 3, the photometric device 1 has five light receiving elements S _{1 to} S ₅ made of SPD (silicon photo diode) or the like corresponding to each of these regions. The photometric device 1 outputs photometric outputs O _{1 to} O ₅ corresponding to the brightness values of the respective areas of the subject screen.

In FIG. 2 and FIG. 3, 8 is an arithmetic processing unit, to which the photometric outputs O _{1 to} O ₅ from the photometric device 1 converted into digital values are input. The arithmetic processing unit 8 calculates the maximum value O _max of the photometric output and the difference ΔO between the maximum value and the minimum value of the photometric output based on these photometric outputs O _{1 to} O ₅ , and outputs this to the neurocomputer 3. .. The method of arithmetic processing is arbitrary, and examples thereof include a method of arithmetically processing an input digital value by software and a method of obtaining an output value by hardware.

[0013] neuro-computer 3 on the basis of the difference ΔO between the maximum value O _{ma x} and the maximum value and the minimum value of the photometric output obtained photometric output by the arithmetic processor 8, required for the operation in the arithmetic unit 4 described below The brightness value BV 'is output. Details of the neuro computer 3 will be described later.

The arithmetic unit 4 includes a lens diaphragm (not shown).
Of the open aperture value AV ₀ of BV = BV, which is the output value from the neurocomputer 3 and the brightness value BV ′.
A calculation of'-AV ₀ is performed to obtain the brightness value BV in which the aperture value of the lens is corrected. Further, the aperture value AV, the film sensitivity SV, and the shutter speed TV are input to the arithmetic unit 4, and APEX (apex) calculation BV + SV = TV + AV is performed to determine the aperture value and the shutter speed. The shutter mechanism unit 6 and the aperture mechanism unit 7 of the exposure control device 5 control the aperture and the shutter based on the aperture value and the shutter speed determined by the arithmetic unit 4.

Reference numeral 9 denotes a control device which incorporates a microcomputer for controlling the entire camera and other peripheral circuits, and is provided with a memory, a judgment circuit, an arithmetic circuit, a timer and the like. The arithmetic unit 4 described above is also included in the control unit 9.

Next, the details of the neuro computer 3 will be described with reference to FIGS. First, in Figure 4,
It is a figure which shows the model of the neuro computer 3 of a present Example. This model was proposed by Rumelhart et al.
-propagation: Back propagation method) model (hereinafter referred to as BP model). The neuro computer 3 is composed of a large number of units (neurons), and each unit is classified into three layers of an input layer, an intermediate layer and an output layer.
The units are connected in a mesh shape in the direction of input layer → intermediate layer → output layer to form a (neural) network. The bond strength of each unit is determined by learning described later. However, it is assumed that there is no mutual connection between units within each layer.

FIG. 5 is a diagram showing a model of a unit (neuron) located in the K layer (K = 0, 1, ..., 0 is an output layer). K + 1 located on the input side, one layer before the K layer
The inputs from the layers are O _p1 ^{K + 1 to} O _pi ^{K + 1} , and the weighting for each input (which corresponds to the coupling strength of the neuron) is W.
_{If j1} ^{K to} W _ji ^K , the output O _pj ^K from the unit is defined by the following equation.

[Equation 1] Note that f (u) is a sigmoid function, and is a function that takes a continuous value between [0, 1] as shown in FIG.

Next, the principle of the learning algorithm of the BP model shown in FIG. 4 will be described. When a certain pattern P (the pattern here is not limited to an image but is considered to be a vector value having n-dimensional elements) is given to the input layer,
_{Let the} actual output value appearing in the output layer be O _pj ^0, and the desired output value at that time (hereinafter referred to as the teacher signal) t _{p j}
The error function E _pj between the two is defined as

[Equation 2]

Therefore, in order to _train this neural network, the error function E _pj should have a minimum value.
It is sufficient to change the coupling strength of all units by using the well-known least mean square error method.

Now, the change amount _Δp W _ji ^K of the coupling strength from the i-th unit of the K-1 layer to the j-th unit of the K layer when the pattern P is given is defined by the following equation.

[Equation 3] Rewriting equation (3) using equation (0),

[Equation 4] Where δ _pj ^K is called the backward propagation amount of error in the Kth layer,

[Equation 5] Is. If we rewrite equation (2) using equation (4),

[Equation 6] In case of output unit

[Equation 7] Therefore, the backward propagation amount δ _pj ⁰ of the error in the output layer is expressed by the following equation.

[Equation 8] In the case of the intermediate unit, since there is no mutual coupling between units in each layer, the backward propagation amount of the error δ
_pj ^K is expressed by the following equation.

[Equation 9] Expression (7) is a recursive function of δ.

When the amount of change _Δp W _ji ^K in the bond strength is generally formulated, it is expressed by the following equation.

[Equation 10] However, Δ _p W _ji ^K (0) = 0, and n represents the number of times of learning. The second term on the right side of the equation (8) is a term added to reduce the error vibration and accelerate the convergence. From equation (8), the bond strength W _ji ^K is updated as follows.

[Equation 11] Where the sigmoid function fi is

[Equation 12] In this case, since fi ′ = fi (1−fi), the backward propagation amount of error is simplified as the following equation. That is, in the case of the output unit

[Equation 13] For the intermediate unit

[Equation 14] As is clear from the above equation, the calculation of Δ _p W starts from the unit of the output layer and proceeds sequentially to the unit of the intermediate layer. In this way, the learning of the neural network proceeds in the opposite direction to the processing of the input data.

Therefore, learning in the BP model is performed as follows. First, input the learning data into the input layer,
Output the result from the output layer. Next, the coupling strength is changed so as to reduce the resulting error (that is, the difference between the actual output and the teacher signal), and the learning data is input again. This operation is
Learning is completed by repeating until _p W converges. Details of this learning algorithm will be described later.

FIG. 7 is a diagram showing the basic circuit configuration of the neurocomputer 3 of this embodiment. In FIG. 7, 11
Is RAM (random access memory), and this RA
M11 has a bond strength W _ji ^K (j = 1, 2, i = 1, 3,
(K = 1 to 3) are stored in a matrix, and three pages are provided for each layer. 12 is also a RAM, and the coupling strength W _ji when the pattern P is given to the RAM 12
^K of variation ^{_{ΔW ji K (j = 1,2,}} i = 1~3, K = 1~
3) are stored in a matrix, and three pages are provided for each layer. 13 is also RAM, and this RAM1
3 is the backward propagation amount of error δ _pj ^K (j = 1, 2, i = 1
˜3, K = 0 to 3) are stored in a matrix, and
Four pages are provided for each layer. 14 is also a RAM, and the output value O _pj ^K from each unit is stored in this RAM 14.
(J = 1, 2, i = 1 to 3, K = 0 to 3) are stored in a matrix, and four pages are provided for each layer.

Reference numeral 15 is an operational circuit of O _pj ^K , which is a RAM 1
4, O _pj ^K is read, data is exchanged with the RAM 11, and the operation result, O _pj ^K, is returned to the RAM 14. 16 is an arithmetic circuit of [delta] _pj ^K, transmits and receives data to and from the RAM14 reads a [delta] _pj ^K from RAM 13, returns the operation result serving [delta] _pj ^K in RAM 13.
_{Reference numeral} 17 is a calculation circuit for Δ _p W _ji ^K , which is calculated from the RAM 12 by Δ W
_ji ^K is read and data is exchanged with the RAMs 13 and 14, and the operation result ΔW _ji ^K is returned to the RAM 12. 18 is an arithmetic circuit of W _ji ^K, RAM 12 reads the W _ji ^K from RAM 11, transmits and receives data to and from the arithmetic circuit 15, and returns the operation result serving W _ji ^K in RAM 11. A control circuit 19 controls the RAM and the arithmetic circuit as a whole.

In the correspondence between the embodiment and the claims, the photometric elements S _{1 to} S ₅ are photometric sensors and the exposure control device 5 is used.
Is the exposure control means, and the RAMs 11 and 14 and the arithmetic circuit 1
5, the RAMs 12 to 13 and the arithmetic circuits 16 to 18 constitute learning means, and the arithmetic processing unit 8 constitutes conversion means.

Next, the learning process of the BP model of this embodiment will be described with reference to the flow charts of FIGS. FIG. 8 is a flowchart of O _pj ^K calculation, and FIG. 9 is δ _pj ^K.
Calculation flowchart, FIG. 10 is a flowchart of W _pj ^K calculation, and FIG. 11 is a flowchart for learning level determination.

First, the program starts from the flowchart of FIG. 8, and in step S101, the bond strength W _ji ^K in the RAM 11 is initialized to a random value. Then
In step S102, the input value O _pj of the input pattern P
⁴ is stored in the RAM 14. In step S103, the value K is assigned to the variable K, and in step S104 it is determined whether the value of the variable K is 0. If the determination is affirmative, the program shifts to the flowchart shown in FIG.
If the determination is negative, the process proceeds to step S105.

In step S105, the value of the variable K is incremented by one, and in step S106, 0 is assigned to the variable j. In step S107, it is determined whether the value of the variable j is 6, and if the determination is affirmative, the program returns to step S104, and if the determination is negative, the program proceeds to step S108. The value of the variable j is incremented by 1 in step S108, and is incremented by 1 in step S109.
The output value O _pj ^K is calculated by the arithmetic circuit 15 based on the output value O _jk ^{K + 1} on the input side _immediately before and the coupling strength W _jk ^K. After that, the program returns to step S107. Through the above procedure, the arithmetic circuit 15 calculates the output value O _pj ^K of the unit.

Next, the program shifts to the flow chart of FIG. 9 and sets the variable K to -1 in step S201.
Is substituted. In step S202, it is determined whether the value of the variable K is 3, and if the determination is affirmative, the program proceeds to the flowchart shown in FIG. 11, and if the determination is negative, the program proceeds to step S203. Step S204
Then, 0 is assigned to the variable j, and it is determined in step S205 whether the value of the variable j is 6, and if the determination is affirmative, the program proceeds to step S206, that is, the subroutine shown in FIG. 10, and the determination is negative. If so, the process proceeds to step S207.

In step S207, the value of the variable j is incremented by 1, and in step S208 it is determined whether the value of the variable K is 0. If the determination is affirmative, the program proceeds to step S209 and the determination is negative. If so, the process proceeds to step S210. In step S209,
Based on the teacher signal t _pj and the output value O _pj ⁰ of the unit in the output layer, the backward propagation amount δ _pj ⁰ of the error in the output layer.
Is calculated by the arithmetic circuit 16. In step S210, based on the output value O _pj ^K of the unit, the backward propagation amount δ _pj ^K-1 of the error in the layer on the output side ahead and the coupling strength W _kj ^K-1 (n + 1) of the unit in this layer. Then, the backward propagation amount δ _pj ^K of the error in the input layer and the intermediate layer is calculated by the arithmetic circuit 16. After that, all the programs return to step S205. Through the above procedure, the backward propagation amount δ _pj ^K of the error is calculated by the arithmetic circuit 16.

In the subroutine of FIG. 10, the bond strength W _ji ^K is calculated as described above. First, step S
In step 301, 0 is assigned to the variable i, and in step S302, it is determined whether or not the value of the variable i is 6, and if the determination is affirmative, the process returns to the main routine, that is, step S202 in the flowchart of FIG. 9, and the determination is negative. If so, the process proceeds to step S303. In step S303, i
Is incremented by 1, and in step S304, the backward propagation amount δ _pj ^K of the error, the output value O _pj ^{K + 1 of the} previous input layer, and the coupling strength W _ji ^K by the previous learning. Based on (n), the variation amount _Δp W _ji ^K (n + 1) of the coupling strength due to learning is calculated by the arithmetic circuit 17.
In step S305, the post-learning joint strength W _ji ^{K is} calculated based on the change amount _Δp W _ji ^K (n + 1) of the joint strength calculated in step S304 and the joint strength W _ji ^K (n) obtained by the previous learning. (n + 1) is calculated by the arithmetic circuit 18. After that, the program returns to step S302. Through the above procedure, the coupling strength W is calculated by the arithmetic circuits 17 and 18.
_ji ^K (n + 1) is calculated.

Then, the program shifts to the flow chart of FIG. 11, 0 is substituted for the variable j in step S401, it is determined in step S402 whether the value of the variable j is 6, and if the determination is affirmative, the program is stepped. S4
If the determination is negative, the process proceeds to step S403.

In step S403, the value of the variable j is incremented by 1, and in step S404, the teacher signal t
_The above-mentioned error function E _pj is calculated based on _pj and the output value O _pj ⁰ of the output layer. After this, the program returns to step S402.

On the other hand, in step S405, the mean square error E _p is calculated, and in step S406 it is judged whether or not the mean square error E _p is smaller than a predetermined threshold value ε. If the judgment is affirmative, the program ends. If the determination is negative, the process returns to step S103 in the flowchart of FIG. The learning level is determined by the above procedure.

The above flow will be described again in accordance with the learning procedure. First, the coupling strength W _ji ^K in the RAM 11 is initialized to a random value (step S101), and the input value O _pj ⁴ of the pattern P is set in the RAM 14 (step S10).
2). Next, the output value O _pj ^K of the unit is sequentially calculated from the input layer to the output layer (steps S103 to S10).
9).

When the first output value O _pj ⁰ is obtained from the output layer, the backward propagation amount δ _pj ⁰ of the error is calculated from this output value and the teacher signal t _pj (steps S201 to S21).
0). Next, the change amount _Δp W _ji ⁰ (1) of the coupling strength is calculated based on the backward propagation amount δ _pj ⁰ of this error (step S).
301-S304). In addition, it is preset to Δ _p W _ji ⁰ (0) = 0. Then, the change amount _Δp W of the bond strength
_ji Based on ⁰ (1) to calculate the binding strength of the output layer W _ji ⁰ (1) (step S305). By the above procedure, O _pj ⁰ , δ _pj ⁰ , Δ _p W _ji ⁰ (1) and W _ji ⁰ (1) of the output layer are obtained. These values are stored in RAM by updating the initial data.
11 to 14 are stored.

Next, learning of the intermediate layer and the input layer is sequentially executed from the intermediate layer toward the input layer. First, δ of the output layer
Based on _{^{pj 0, W ji 0 (1}} ) and O _pj ¹ in RAM 14, it calculates the backward propagation amount [delta] _pj ¹ error in the intermediate layer (step S201~S210). Next, based on the backward propagation amount δ _pj ¹ of this error, the change amount Δ _p W _ji ¹ (1) of the coupling strength
Is calculated, (step S301 to S304), based on the amount of change delta _p W _ji ¹ of this bond strength (1) to calculate the binding strength of the output layer W _ji ¹ (1) (step S305). By the above procedure, δ _pj ¹ , Δ _p W _ji ¹ (1) and W _ji ¹ of the intermediate layer
(1) is obtained. These values are stored in the RAMs 11 to 14 by updating the initial data. Then, the above procedure is sequentially repeated toward the input layer, and the first learning ends.

Learning is executed a plurality of times (n) times according to the procedure described above, and the coupling strength W _ji ^K between each unit is determined.
Then, every time learning is completed, the error function E _pj and the mean square error E _p are calculated (steps S401 to S40).
5) When the mean square error E _p becomes smaller than the threshold value ε, the learning ends.

In the present embodiment, the pattern P given to the neurocomputer 3 is a subject screen such as a typical landscape or person, and the teacher signal is a luminance value desired by the photographer for this subject screen. It is BV 'itself. The brightness value BV ′, which is a teacher signal, is given to the neurocomputer 3 by a dial switch or the like (not shown). At this time, a plurality of patterns P may be given, and it is possible to perform learning so as to obtain different output values (brightness value BV ′) for each pattern P.

Neurocomputer 3 for which learning has been completed
When a pattern P (output from the arithmetic processing unit 8) is input to the input layer, the arithmetic circuit 15 shifts from the input layer to the output layer based on the coupling strength W _ji ^K between the units determined by learning. Then, the output value O _pj ^K is sequentially calculated, and the output value O _pj ⁰ (luminance value BV ′) in the output layer is output to the outside as the output value of the neurocomputer 3. The brightness value BV ′ output from the neurocomputer 3 is then input to the arithmetic unit 4 to perform APEX calculation to determine the aperture value and shutter speed, and the shutter mechanism unit 6 and the aperture mechanism unit 7 are operated by the arithmetic unit. The aperture and the shutter are controlled based on the aperture value and the shutter speed determined in 4.

Therefore, according to the present embodiment, it is possible to obtain the desired exposure result by performing the desired exposure control desired by the photographer by learning the neurocomputer 3. Moreover, in the present embodiment, the photometric outputs to be input to the neurocomputer 3 are collected into two types of variables by the arithmetic processing unit 8. Therefore, the scale of the network in the neurocomputer 3 can be small, and the learning time can be shortened compared with the conventional one. And the calculation time can be shortened. In addition, the storage capacity of the storage medium (RAM 11 to 14) can be reduced as the number of variables decreases, and it becomes easy to store this storage medium in the limited space in the camera. Further, by removing the variables that do not directly contribute to the output of the network by the arithmetic processing unit 8, the learning time and the arithmetic time can be shortened, and the situation that the learning does not converge or the correct value is not output is prevented. be able to.

The above discussion will be further described with reference to FIG. FIG. 12 is a diagram showing the scale of the RAM 11 shown in FIG. As shown in FIG. 12A, when the number of input variables is 5, that is, when the photometric outputs from the light receiving elements S _{1 to} S ₅ are directly input, the matrix of W _ji ^K is 5 for each layer. Although x25 = 125 elements are required, as shown in FIG. 9B, when there are two input variables as in this embodiment, the matrix of W _ji ^K is 2 × 6 = 12 elements for one layer. Is enough. In this way, the scale required for the RAM 11 is reduced in geometric progression only by slightly reducing the input variables.
The above discussion is the same for the other RAMs 12 to 14, and the network is dramatically simplified.

Second Embodiment FIG. 13 shows a second camera having a learning function according to the present invention.
It is a figure which shows an Example. In this embodiment, the arithmetic processing unit 8
The other components are the same as those in the above-described first embodiment except for the configuration described above, and therefore, illustration and description of the other components are omitted.

The arithmetic processing unit 8 of the present embodiment calculates the photometric center O ₁ , the peripheral sky O ₂ + O ₃ , and the peripheral area O ₄ + O ₅ with respect to the input photometric outputs O _{1 to} O ₅ . Then, this is output to the neuro computer 3. Therefore, according to this embodiment as well, it is possible to obtain the same effects as those of the first embodiment described above.

[Third Embodiment] FIG. 14 shows a third camera having a learning function according to the present invention.
It is a figure which shows an Example. In this embodiment, the light receiving element S ₁ to S _n constituting the photometric apparatus 1 are arrayed in a matrix, are input to the photometry output O ₁ ~ O _n all processor 8 of the light-receiving elements There is. Processor 8 calculates the difference ΔO between the maximum value O _max and the maximum value and the minimum value of the photometry output of those photometric output O ₁ ~ O _n outputs the neuro-computer 3. Alternatively, the subject screen is divided into five regions, that is, the central part, the upper left part, the upper right part, the lower left part and the lower right part, and a large number of light receiving elements are associated with these regions, and the photometric central part O ₁ and the peripheral part O ₂ + O ₃ , the peripheral area O ₄ + O ₅ is calculated and output to the neurocomputer 3.

Therefore, also in this embodiment, the above-mentioned first
It is possible to obtain the same effect as that of the embodiment. That is, according to this embodiment, even if the number of light receiving elements increases, the number of variables input to the neurocomputer 3 by the arithmetic processing unit 8 can be made constant, and the storage capacity and the calculation amount increase as the number of sensors increases. Can be suppressed.

The details of the camera having the learning function of the present invention are not limited to the above-mentioned embodiments, and various modifications are possible.

[0048]

As described in detail above, according to the present invention, the parameters to be input to the network are summarized by the converting means into the parameters whose number is smaller than the output signal from the photometric sensor. The size is small, and the learning time and the calculation time can be shortened as compared with the related art. Further, as the number of variables decreases, the storage capacity of the storage medium also becomes small, and it becomes easy to store this storage medium in the limited space in the camera. Furthermore, by removing the variables that do not directly contribute to the output of the network by the conversion means, it is possible to shorten the learning time / calculation time and prevent the situation where the learning does not converge or the correct value is not output. it can.

[Brief description of drawings]

FIG. 1 is a diagram corresponding to a claim of the present invention.

FIG. 2 is a block diagram showing a circuit configuration of a camera having a learning function according to the first embodiment of the present invention.

FIG. 3 is a diagram showing details of an arithmetic processing unit and a light receiving element.

FIG. 4 is a diagram showing a model of a neuro computer.

FIG. 5 is a diagram showing a model of a unit.

FIG. 6 is a diagram showing an example of a sigmoid function.

FIG. 7 is a block diagram showing a basic circuit configuration of a neurocomputer.

FIG. 8 is a flowchart for explaining a learning process of a neuro computer.

FIG. 9 is a view similar to FIG.

FIG. 10 is a view similar to FIG.

FIG. 11 is a view similar to FIG.

FIG. 12 is a diagram for explaining the effect of the present invention.

FIG. 13 is a diagram showing a camera having a learning function according to a second embodiment of the present invention, which is a diagram showing details of an arithmetic processing unit and a light receiving element.

FIG. 14 is a diagram showing a camera having a learning function according to a third embodiment of the present invention, which is a diagram showing details of an arithmetic processing unit and a light receiving element.

[Explanation of symbols]

a photometric sensor b exposure control means c network d learning means e conversion means 1 photometric device 3 neurocomputer 4 arithmetic device 5 exposure control device 6 shutter mechanism unit 7 diaphragm mechanism unit 11-14 RAM 15-18 arithmetic circuit S ₁ -S _n the light-receiving element O ₁ ~O _n photometric output

Claims (1)

Claim: What is claimed is: 1. A plurality of photometric sensors that output photometric information, an exposure control unit that controls the exposure of a camera based on output signals of the photometric sensors, and an output signal of the photometric sensors. A network capable of learning by self-organization, which sends an exposure control signal to the exposure control means as a parameter, and a learning means for self-organizing the network so that the output value of the network matches a desired output value. In a camera having a learning function, provided is a conversion unit interposed between the photometric sensor and the network for converting a plurality of output signals from the photometric sensor into a smaller number of network input parameters. A camera with a learning function.