JPH04330411A - Automatic focusing device - Google Patents

Abstract

PURPOSE:To predict a focusing position without necessitating special conditions in order to predict the focusing position and to predict the focusing position by using a neural network the case of detecting the focusing position in order to accurately detect the focusing position. CONSTITUTION:The signal of an image picked up by a photographing optical system 2 and an image pickup element 3 is read out and a focusing signal value in accordance with focusing degree is detected by a focusing signal detector 32. Meanwhile, the focusing position is predicted based on the learning of a focusing position detector 33 provided with the neural network from the positions of the plural photographing optical systems 2 and the plural focusing signals corresponding to the positions thereof. According to focusing position information outputted from the detector 33, a motor control circuit 27 controls a motor 17, so that the relative position between the optical system 2 and the image pickup element 3 is changed along an optical axis direction.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] This invention relates to an automatic focusing device.
In particular, it relates to an automatic focusing device that is used in single-lens reflex cameras, electronic cameras, etc., and uses a neural network for focus detection.

0002

[Prior Art] A conventionally known automatic focusing device extracts a predetermined frequency component from an image signal obtained from an image sensor, and moves the photographing optical system to a position where the extracted frequency component is maximum. There is a so-called hill-climbing method in which focus is adjusted by

[0003] Automatic focusing devices that apply this hill-climbing method do not require special optical parts for focus adjustment, so they can be made smaller, and they can achieve high-precision focusing without depending on the pattern of the subject. It has the advantage of being possible. An automatic focusing device using the mountain climbing method is available from NHK, for example.
Technical Report, 1965, Volume 17, No. 1, Volume 86 (
P21 to P37).

[0004] Figure 11 shows the configuration of the conventional mountain climbing method.
Figure 2 shows a characteristic diagram of its operation. A light flux from a subject 1 passes through a photographic optical system 2, forms an image on an image sensor 3, is converted into an electrical signal, and is then inputted to a camera circuit 5 for camera processing via an amplifier 4. At the same time, it is also input to a band pass filter (hereinafter abbreviated as BPF) 6, and only a predetermined frequency component of the video signal is extracted.
is input.

Furthermore, from the horizontal synchronizing signal HD and vertical synchronizing signal VD separated by the camera circuit 5, a window pulse forming circuit (window) 8 forms a so-called window pulse for extracting only a predetermined area of the screen. This window pulse is applied to the gate circuit 7.
is input, and the signal of the above-mentioned predetermined frequency component is extracted only during this pulse generation period. As the extracted output, a signal corresponding to the degree of focus (hereinafter referred to as a focus signal) is outputted by a detector (DET) 9 and an integrator 10. Here, the position of the photographing optical system 2 where the focus signal is maximum corresponds to the in-focus position.

Therefore, focus adjustment is performed by controlling the position of the photographing optical system 2 so as to climb the peak of the focus signal as shown in FIG. This control is performed using the sample hold circuit 11 and one field delay.
This is done by a hill climbing circuit 14 consisting of a circuit 12 and a comparator 13. Sample hold is mono multivibrator (MM) 15, sample pulse forming circuit (S.P)
16, this is performed for each field using sample pulses generated at a constant timing. And integrator 1
The focus signal, which is the output of 0, is held by the sample and hold circuit 11 for each field, and the comparator 13 compares the previously held value output from the 1-field delay circuit 12 with the current held value for each field. The photographing optical system 2 is driven in the direction in which the signal increases. Furthermore, 1
7 is a motor for driving the photographing optical system 2, and 18 is a driving circuit thereof.

Now, as shown in FIG. 12, if the focus signal is A in the previous field and B in the next field, then A<B, so the driving direction of the photographing optical system 2 is maintained as it is. Furthermore, if the focus signal is C in the next detection, since B<C, the driving direction of the photographing optical system 2 is still maintained as it is. Then, in the next detection, the focus signal becomes D
If so, C>D, and it is detected that the in-focus position has been passed, and the driving direction of the photographing optical system 2 is reversed. Then, the photographing optical system 2 is driven again toward the focusing direction. After that, this operation is repeated, and finally, the focus is adjusted by reaching a steady state while vibrating little by little near the top of the mountain.

[0008]

However, in the conventional mountain-climbing method described above, the object must pass past the in-focus position once, regardless of whether it is in focus or not. Therefore, not only does the time required for focusing become longer,
Since the photographing optical system 2 vibrates little by little, the photographer may feel uncomfortable during the focusing operation.

[0009] As a conventional example to eliminate such drawbacks,
For example, there is a technique described in Japanese Unexamined Patent Publication No. 62-208015. This will be explained with reference to FIGS. 13 and 14.

In FIG. 13, reference numeral 19 denotes a one-field delay circuit, which receives the output of the one-field delay circuit 12 as an input. In other words, the focus signal C of the current field, the focus signal B of the previous field, and the focus signal A of the field before the previous field.
is being output. In addition, 20 and 21 are subtracters, 22 and 2
3, 24, 25 are comparators, 26 are comparators 22, 23, 2
4 is an AND circuit with an output, and 27 is a motor control circuit. Furthermore, 28 is a summit prediction circuit, and 29 is a photographing optical system 2.
A lens information sensor 30 is a control circuit that controls the BPF 6, the gate circuit 7, and the top prediction circuit 28.

The above-mentioned summit prediction circuit 28 calculates the following equations 1 and 2.
The top of the mountain is predicted when the following conditions are satisfied:

0012

[Math 1]

[0013]

[Math 2]

[0014]

[Math 3]

[0015] Here, Vref I and Vref II are predetermined constants.

In other words, when the focus signal of the current field is sufficiently large (C>Vref I) and the difference between the focus signals of the current field and the previous field is below a predetermined level and sufficiently close to the peak (C-B<Vref II). ), and when the slope of the current mountain is smaller than the slope of the previous mountain (C-B<B-A), the summit prediction circuit is activated and the focus signal P of that mountain is calculated. Use a simple quadratic function to predict how many more fields it will take to reach the top. After the predicted number of fields (N), the focus signal C becomes P≧
In case C, the motor is stopped and focus adjustment is completed. still,
The quadratic function used for prediction is expressed by the formula y=ax2+bx+c, where x is the number of fields and y is the focal signal.

However, in the prior art, the above equations 1 and 2
and the condition of Equation 3 is often not satisfied. For example, in the case of a low-contrast subject, the focus signal value is small even at the in-focus position, and the condition of Equation 1 cannot be satisfied. Further, if Vref I is lowered to avoid this, the condition of Equation 1 will be satisfied at a position quite far from the in-focus position for a subject with high contrast, and the prediction accuracy of the top will deteriorate.

[0018] Also, it is explained that a quadratic function is used to predict the summit, but as described in "Introduction to Optics II - Wave Optics -" (written by Junpei Tsujiuchi, Asakura Shoten, pp. 120-121), OT when the diameter of the circle of confusion at defocus is 2b
F (Optical Transfer Function
n) is as shown in equation 4.

[0019]

[Math 4]

Here, s is a spatial frequency, and J1 is a first-order Bessel function of the first kind. Further, this equation has the same change for b, and since b is proportional to the defocus amount x, if 2b=x, Equation 4 can be expressed as Equation 5 below.

[0021]

[Math 5]

This corresponds to the curve formed by the focal signal in the case of a narrow band. In reality, the BPF 6 is not a narrow band but has a certain width, and if this characteristic is b(s), the image spectrum is F(s), and the detector 9 is a squarer, then the integrator 1
The focus signal output from 0 is as shown in Equation 6.

[0023]

[Math 6]

Here, it can be seen that the shape of the curve of the focal signal differs depending on the spectrum F(s) of the image, and that prediction of the peak by simple approximation of a quadratic function is theoretically incorrect and has insufficient accuracy.

[0025] Furthermore, if the photographing optical system is near the in-focus position at the start of focus adjustment, the condition of Equation 1 is not satisfied;
In the end, the conventional hill-climbing method is used, and the above-mentioned problem of small vibrations near the focusing position cannot be solved.

[0026] As described above, the present invention, which was conventionally difficult to call practical for the purpose of predicting the in-focus position, was made in view of the above problems, and a special It is an object of the present invention to provide an automatic focusing device capable of predicting a focus position without requiring conditions and detecting the focus position with high accuracy.

[0027]

[Means for Solving the Problems] That is, the present invention provides a photographic optical system, an image sensor for capturing an image formed by the photographic optical system, and a relative position between the image sensor and the photographic optical system on an optical axis. a driving means for changing the electric charge along the direction; an image reading means for reading out the charge accumulated in the image sensor as an image signal; and a bandpass filter for passing only a predetermined frequency band from the image signal read by the reading means. , a focus signal detection means for detecting a focus signal value according to the degree of focus from the output signal of this bandpass filter, a focus position information detection means having a neural network, and this focus position information detection means It is characterized by comprising a control means for controlling the driving means according to focus position information output from the drive means.

[0028]

[Operation] The automatic focusing device of the present invention utilizes a neural network trained to predict a focus position from the positions of a plurality of photographic optical systems and a plurality of focus signals corresponding thereto. By using a neural network as a means to detect focus position information and driving and controlling the photographing optical system according to the output focus position information, special conditions are not required for prediction. Therefore, it is possible to provide an automatic focusing device that can predict the in-focus position and detect the in-focus position with high accuracy.

[0029]

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings.

First, a neural system which is applied to the automatic focusing device of the present invention and which is learned to predict the focusing position from the positions of a plurality of photographing optical systems and a plurality of focus signals corresponding thereto. Describe networks.

FIG. 4 is a conceptual diagram of this neural network. In the figure, 31 is a neural network, which includes an input layer 31a, an intermediate layer 31b, and an output layer 31.
It consists of three layers of c. The input is the position of the photographing optical system
, X2, X3, and the focus signal P corresponding to these
1, P2, and P3. Here, P1 is the most recently detected focus signal, P2 is the previous focus signal of P1, P3
represents the previously detected focus signal of P2. The output Y is learned to give the distance from the photographing optical system position X1 to the in-focus position. This learning is performed using a backpropagation algorithm using the position of the photographing optical system and the focus signal as learning data, and the difference α between the in-focus position and the position of the photographing optical system as a teacher signal. The weight of each neuron unit is designed so that the multiplicative error is minimized.

Next, a first embodiment of the automatic focusing device of the present invention will be described.

FIG. 1 is a block diagram of an automatic focusing device. In the figure, the light flux from subject 1 is transmitted to photographing optical system 2.
After passing through the amplifier 4 and forming an image on the image sensor 3 and converting it into an electrical signal, the focus signal P is detected by the focus signal detector 32 via the amplifier 4.

The focus signal detector 32 is constructed as shown in FIG. In the example of the focus signal detector 32 shown in FIG. 2(a), after the image signal is input to the BPF 6 having the passing center spatial frequency ST, only the signal at the image position specified by the ranging area specifying circuit 35 is gated. The focus signal P of the spatial frequency ST is extracted by the detector 9, the A/D converter 36, and the digital integration circuit 37. Note that the digital integration circuit 37 is composed of an adder 38 and a latch circuit 39.

Focus signal detector 32 shown in FIG. 2(b)
In the example of ', first, the A/D converter 36 converts the signal into a digital signal, and then the digital BPF 40 extracts the frequency component of the spatial frequency ST.
Only the signal at the image position specified by 5 is sent to the gate circuit 7.
Extracted by Here, the root mean square value and the average value are 2
The root mean value detection circuit 41 and the mean value detection circuit 42 respectively obtain the dispersion value, and the dispersion detection circuit 43 obtains the dispersion value shown in Equation 7 as the focus signal P.

[0036]

[Math 7]

Since σ is a variance, it is not affected by changes in bias, and since it is normalized by the average value, it is also not affected by changes in gain. Therefore, even if there is flicker or a sudden change in illumination light, it can be used as an accurate focus signal. Note that the BPF 6 or the digital BPF 40 may be one whose passing center frequency can be freely varied. Further, the passing center frequency of the BPF 6 is set by a control circuit (not shown), and the passing center frequency (hereinafter referred to as band information) s is input to the focus position detector 33 of FIG.

Returning to FIG. 1, the focus signal P output from the focus signal detector 32 is input to the focus position detector 33, and is compared with the position X of the photographing optical system 2 obtained from the photographing optical system information sensor 29a. The distance Y to the in-focus position is determined from the aperture value F and the band information s. The motor control circuit 27 performs focus adjustment by driving the motor 17 from a distance Y to the focus position. Furthermore, 34 is a learning device, and a focus position detector 33
The weight between each neuron within the focus position detector 33 (details will be described later) is set from the distance Y to the focus position output from the focus position detector 33 and the teacher signal α. This learning device 34
As shown in FIG. 3, a subtracter 60, a squarer 61,
It is composed of an adder 62 and a weighting coefficient generator 63.

The focus position detector 33 is constructed as shown in FIG. As described above, the neural network 31 includes an input layer 31a, an intermediate layer 31b made up of nonlinear neuron units, and an output layer 31c made up of linear neuron units. input layer 31
There are (m+n) pieces of a, q pieces of intermediate layer 31b, and output layer 31c.
consists of one unit. Here, it is assumed that q≦m+n. Furthermore, there is no connection between neuron units belonging to the same layer, there is only connection between neuron units in adjacent layers, and learning is performed under the conditions of aperture value FT and BPF band ST.

A detailed diagram of each neuron unit is shown in FIG. In FIG. 6, neuron units 31b-j (
j=1 to q) executes the calculation shown in Equation 8.

[0041]

[Math. 8]

In other words, the inputs I1 to Im+n are the weighting coefficient values Wji to Wj(n+n) output from the weighting coefficient memories 54-j1 to 54-j(m+n)) to the multipliers 55-j1 to 55-j( m+n)), and the sum thereof is calculated by an adder 56j. From this sum,
After the threshold value θj is subtracted, the nonlinear converter 59j performs the conversion shown in Equation 9.

[0043]

[Math. 9]

Here, Equation 9 is called a sigmoid function. Note that the weighting coefficient memories 54-j1 to 54-j (m+
n) and the threshold memory 57j are set by the weight signal Ws output from the learning device 34 in FIG.

Furthermore, in this embodiment, in order to change the weighting coefficient and threshold value of each neuron according to the number of inputs to the input layer as described later, the weighting coefficient and threshold value for the number of inputs r are , θj r . Each weighting coefficient memory and threshold value memory is configured to store a maximum input number N values.

Further, the neuron unit 31c in FIG. 5 has almost the same configuration as the neuron unit in FIG. 6, but it is a linear converter instead of a nonlinear converter. The output width can be expanded by using a linear converter.

The focus signal P input from the focus detector 32 of FIG. 1 every field time is sequentially stored in the focus signal storage circuit 46 shown in FIG. Then, they are inputted to the focus signal preprocessing circuit 44 as P1, P2, P3, ... in the newly detected order, converted into m pieces of data by the focus signal preprocessing circuit 44, and inputted into the neural network 31 as I1 to Im. Input layer 31a-1 to 31a
-m is input. On the other hand, the positions of the photographing optical system when the focus signals are detected are sequentially stored in the photographing optical system position storage circuit 47, and the detection positions X1, X2, X3, ... of the focus signals P1, P2, P3, ... are photographed. The signal is input to the optical system position preprocessing circuit 45. Then, it is converted into n pieces of data having information on the position of the photographing optical system 2, and is sent to the input layer 31a-(m+1) to 3 of the neural network 31 as Im+1 to Im+n.
1a-(m+n).

The inputs I1 to Im+n pass through the intermediate layer 31b and the output layer 31c, and are outputted from the output layer 31c as the distance YT from the photographing optical system position to the focal position in the case of the aperture value FT and the BPF band ST. Ru. This distance YT is divided by the width normalization coefficient Rc (described later) output from the photographic optical system position preprocessing circuit 45 in the divider 54, and the distance Y (=YT/Rc) to the in-focus position is obtained. .

The focus signal preprocessing circuit 44 shown in FIG. 5 normalizes the magnitude of the focus signal, which varies depending on the subject, and is configured as shown in FIG. 7(a) or (b).

The example of the focus signal preprocessing circuit 44 shown in FIG. 7A has a maximum value detector 47 for the focus signal P, and a divider 48 for each focus signal P by the detected maximum value Pmax.
Normalization is performed by dividing by -1 to 48-N, and I1
~IN is output. In this case, the number of inputs is m=N.

On the other hand, the example of the focus signal preprocessing circuit 44' shown in FIG.
~49-(N-1), and divider 50-
Normalization is performed by dividing by the average value of the image signal of the subject at 1 ~ 50 - (N-1), and I1 ~
Output IN-1. In this case, the number of inputs is m=N-1.

Returning to FIG. 5, the photographing optical system position preprocessing circuit 47 normalizes the steepness of the focus signal, which varies depending on the aperture value F and the BPF passing center frequency s. As shown in FIG. 8, it is known that the horizontal width of the curve formed by the focus signal is proportional to the aperture value and inversely proportional to the BPF passing center frequency s (see Japanese Patent Application No. 2-303285). For example, if the aperture value is doubled or the BPF band is halved,
The width of the curve formed by the focus signal is doubled, and when the aperture value is 1/2 times or the BPF band is doubled, the width of the curve formed by the focus signal becomes 1/2 times.

Therefore, the focus signal can be normalized according to the aperture value and the BPF passing center frequency, but since the position of the photographing optical system 2 itself cannot be normalized unless the in-focus position is known, the position of the photographing optical system 2 cannot be normalized. Normalize the difference in position. In other words, the difference d between the two positions in the focus signal detected at the aperture value F and the BPF passing center frequency s is calculated as the difference d' between the two positions when the aperture value FT and the BPF passing center frequency ST are used. is converted to

[0054]

[Math. 10]

For example, in FIG. 8, the difference d2 between the two positions when the aperture value is 2FT is converted to the difference d0 between the two positions when the aperture value is FT. Here, we define equation 11 and Rc
is called the width normalization coefficient.

[0056]

[Math. 11]

FIG. 9 is a configuration diagram of the photographing optical system position preprocessing circuit 46. As shown in FIG. 51-1 to 51-(N-1) are subtracters, which detect the difference in the position of the photographing optical system 2, and calculate the width normalization coefficient R obtained by the dividers 55, 56 and the multiplier 57.
By multiplying c by multipliers 53-1 to 53-(N-1), normalized data Im+1 to Im
+(N-1) is output. In this example, the number n of inputs in the input layer is n=N-1. The width normalization coefficient Rc is also output and used to return the normalized YT to the actual aperture value F and BPF passing center frequency s.

Next, the operation of this embodiment will be explained. Note that there are two modes of operation: a learning mode and an execution mode.

First, the learning mode will be explained. Learning is performed using the backblopage algorithm (D. Rumelhart et al.).
E. Rumelhart, J. L. McClelland, PDP Research Group, translated by Shunichi Amari, "PDP Model" Chapter 8, Sangyo Tosho, 1989), and the weighting coefficient of each neuron unit is set.

First, set the aperture value of the photographing optical system 2 to FT,
Set the BPF band of the focus signal detector to ST. Then, the focus signal P is detected while changing the distance U to the subject and the position X of the photographing optical system 2, and P and
3, and the learning device 34 starts learning. here,
The accurate distance to the subject is measured by another means, and the focal position Xf of the photographic optical system 2 is detected.

The teacher signal α is obtained from this Xf and the current photographing optical system position X1, and when r<N, learning is performed using the equation 12
This is done by changing the weight and threshold of each neuron unit so as to minimize the .

[0062]

[Math. 12]

[0063] Here, X represents the position of the photographing optical system, from the position for photographing an infinite object (X∞) to the position for photographing a subject according to the number of inputs (X∞+(r-1)dx). ,
The position of the photographing optical system incorporated in the photographing optical system 2 is changed every predetermined step (dx).

[0064] If this predetermined step is made larger than the minimum step, the number of times of learning will be reduced, leading to an improvement in learning speed. In addition, U is the distance to the subject, and from an infinitely far subject to a very close subject, for example, multiple subjects at distances such that the imaging distance is equally spaced (U∞, U1,
U2,..., Unearest). Yux is the output of the neural network when the object distance is U and the photographic optical system position is X, and αux is the teacher signal in that case.

Furthermore, the focus signal storage circuit 46 outputs the number r of input focus signals as an input number signal to each neuron unit, and the weighting coefficient Wjir and threshold value θj r are selected according to this r. Equation 12 above indicates that learning is performed for each r.

[0066] When the number of inputs is one, it is almost impossible to detect the in-focus position, so when the number of inputs is two or more, learning is performed as described in I and II below.

I. When the number of inputs r=2 (1) Each weight parameter Wji2, threshold value θj
Initialize 2 with a random number. (2) Clear adder 62 to zero. (3) Drive the photographing optical system 2 to the position of X=X∞, set the subject to the position of U=U∞, and start detecting the first focus signal.

(4) The photographing optical system 2 is driven by a predetermined step (X=X∞+dx), and the second focus signal is detected. Since the number of inputs r=2, the focus signal P1,
P2 is input to the focus signal preprocessing circuit 44. At this time, P3 to PN are input as zero. Similarly, the photographing optical system positions X1 and X2 are input to the photographing optical system position preprocessing circuit 45, but X3 to XN are input as zero. Then, the normalized values I1 and I2 of the focus signal and the normalized value Im+1 of the photographing optical system position are input to the neural network 31. Here, zero is input to I3 to Im and Im+2 to Im+n. Further, the output YT from the neural network 31 is divided by the width normalization coefficient Rc, and Equation 13 is detected and input to the learning device 34.

[0069]

[Math. 13]

In this learning device 34, the subtracter 60 calculates the number 14.
A difference operation with the teacher signal shown in is performed, and the difference is calculated by the squarer 6
The signals are squared by 1 and added by an adder 62.

[0071]

[Math. 14]

(5) The above processes (3) to (4) are performed with the distance to the object being U=U1. From now on, U=U
2, U3, ..., until Unearest (3) ~
Repeat the process in (4) to detect E in number 12.

(6) Next, change the weight of each neuron unit so that the value of E decreases. Here, the i-th weight change amount ΔWjir of the neuron unit j that causes E to decrease the most can be found using Equation 15 using the steepest descent method. Note that the symbol ε in the formula is a positive number here.

[0074]

[Math. 15]

(7) Thereafter, processes (2) to (6) are repeated, and the process is terminated when E in Equation 12 becomes sufficiently small. Then, the weight values at that time are stored in the weight coefficient memory 54 and the threshold value memory 5 of each neuron unit.
7 as a weighting coefficient Wji2 and a threshold value θj2.

II. When the number of inputs r=3 (1) Each weight parameter Wji3 and threshold value θj3 are initialized with random numbers. (2) Clear adder 62 to zero. (3) Drive the photographing optical system 2 to the position of X=X∞, set the subject to the position of U=U∞, and start detecting the first focus signal.

(4) The photographing optical system 2 is driven by a predetermined step (X=X∞+2dx), and the second focus signal is detected.

(5) The photographing optical system 2 is further driven by a predetermined step (X=X∞+dx), and the third focus signal is detected. Since the number of inputs r=3, the focus signal P
1 , P2 , and P3 are input to the focus signal preprocessing circuit 44 . At this time, P4 to PN are input as zero. Similarly, the photographing optical system positions X1, X2, and X3 are input to the photographing optical system position preprocessing circuit 45, but X4 to XN
is entered as zero. Then, the normalized values I1, I2, I3 of the focus signal and the normalized values Im+1, Im+2 of the photographing optical system position are input to the neural network 31. Here, I4 ~Im, Im+3 ~Im
+n is input with zero. Further, the output YT from the neural network 31 is divided by the width normalization coefficient Rc, and Equation 16 is detected and input to the learning device 34.

[0079]

[Math. 16]

In this learning device 34, the subtracter 60 calculates the number 17.
A difference operation with the teacher signal shown in is performed, and the difference is calculated by the squarer 6
The signals are squared by 1 and added by an adder 62.

[0081]

[Math. 17]

(6) Processes (3) to (5) are performed with the distance to the object being U=U1. From now on, U=U2
, U3 ,..., until Unearest (3) ~ (5
) is repeated to detect E in number 12.

(7) Next, the weight of each neuron unit is changed based on Equation 15 so that the value of E decreases.

(8) Thereafter, processes (2) to (7) are repeated, and the process is terminated when E in Equation 12 becomes sufficiently small. Then, the weight values at that time are stored in the weight coefficient memory 54 and the threshold value memory 5 of each neuron unit.
7 as a weighting coefficient Wji3 and a threshold value θj3.

The above steps are repeated until r=4 to N-1, and the weighting coefficients Wji4 to WjiN-1 and the threshold value θj are
4 ~ Find the threshold value θj N-1. The weighting coefficient WjiN and the threshold value θj N are determined as follows.

(1) Each weight parameter WjiN,
Initialize the threshold value θj N with a random number. (2) Clear adder 62 to zero. (3) Drive the photographing optical system 2 to the position of X=X∞, set the subject to the position of U=U∞, and start detecting the first focus signal.

(4) The photographing optical system 2 is driven by a predetermined step (X=X∞+dx), and the second focus signal is detected.

(5) Thereafter, the process in (4) is repeated to detect the focus signal at the position X=X∞+(N-1)dx. Since the number of inputs r=N, the focus signal P1. P
2, ~PN are input to the focus signal preprocessing circuit 44. Similarly, the photographing optical system positions X1, X2, to XN are input to the photographing optical system position preprocessing circuit 45. Then, the normalized values I1, I2, ~Im of the focus signal and the normalized values Im+1, Im+2, ~Im+n of the photographing optical system position are input to the neural network 31. Furthermore, the output YT from the neural network 31 is divided by the width normalization coefficient Rc to obtain the equation 18.
is detected and input to the learning device 34.

[0089]

[Math. 18]

In this learning device 34, the subtracter 60 calculates the number 19
A difference operation with the teacher signal shown in is performed, and the difference is calculated by the squarer 6
The signals are squared by 1 and added by an adder 62.

[0091]

[Math. 19]

(6) Perform the processes in (3) to (5) with the distance to the object being U=U1. From now on, U=U2
, U3 ,..., until Unearest (3) ~ (
Repeat the process in 5).

(7) Next, set the photographing optical system 2 to X=X∞
It is driven to the +dx position and processes (2) to (6) are performed. Here, in the process of (5), X=X∞+dx
Perform at position +(N-1)dx. In this way, (5)
The process ends until the position X=Xnearest+(N-1)dx, and E is detected. This E can be expressed as the number 20.

[0094]

[Math. 20]

(8) Next, the weight of each neuron unit is changed based on Equation 20 so that the value of E decreases.

(9) Thereafter, processes (2) to (8) are repeated, and the process is terminated when E in Equation 12 becomes sufficiently small. Then, the weight values at that time are stored in the weight coefficient memory 54 and the threshold value memory 5 of each neuron unit.
7 as a weighting coefficient WjiN and a threshold value θj N .

[0097] In this way, all weighting coefficients and threshold values are determined and the learning mode ends. Here, the accuracy may be further improved by changing the type of subject.

Next, the execution mode will be explained. In this execution mode, the learning device 34 is not required. When focus adjustment begins, the photographing optical system 2 is first driven to a position where X=X∞. Then, by an exposure detector (not shown),
An appropriate aperture value F is set according to the brightness of the subject. At this time, the exposure time of the image sensor 3 may also be changed appropriately. Then, at the same time as detection of the focus signal of the passing center spatial frequency ST starts, the photographing optical system 2 starts to be driven at a constant speed in the direction of the position (Xnearest) for photographing the nearest object. However, this speed is such that the detection of the focus signal is a step smaller than a predetermined step during learning.

When the second focus signal is detected, the focus signals P1 and P2 are input to the focus signal detection circuit 44, the focus signal values are normalized, and the focus signals P1 and P2 are input as inputs I1 and I2. It is input to the input layers 31a-1 to 31a-2 of the neural network. in this case,
Zero is input to the input layers 31a-3 to 31a-m.

On the other hand, the photographing optical system positions X1 and X2 corresponding to the focus signal are input to the photographing optical system position preprocessing circuit 45, normalized by the width normalization coefficient Rc, and then input to the input I
input layer 31a of the neural network as m+1
-(m+1) is input. In this case, the input layer 31a
Zero is input to -m+2 to 31a-m+n. Then, in the neuron unit 31, the input signal r(r
= 2), the weighting coefficient Wji2 in the neuron unit
A threshold value θj 2 is selected.

[0101] In this way, the aperture value FT is obtained from the output layer 31c.
, the distance YT to the in-focus position in the case of the band ST is determined and divided by Rc in the divider 54, and the distance Y to the in-focus position is detected. Then, the motor control circuit 27 drives the motor 17 based on this value of Y, and the photographing optical system 2
is moved to the in-focus position.

Then, when the focus signal of the third point is detected, the focus signals P1, P2, P3 are input to the focus signal detection circuit 44, the focus signal value is normalized, and the input I1, They are input as I2 and I3 to the input layers 31a-1 to 31a-3 of the neural network. At this time, zero is input to the input layers 31a-4 to 31a-m. On the other hand, the photographing optical system positions X1, X2, and X3 corresponding to the focus signals are input to the photographing optical system position preprocessing circuit 45, and after being normalized by the width normalization coefficient Rc, they are input to the neural network as inputs Im+1 to Im+2. Input layer 31a-(m+1) ~31
It is input to a-(m+2).

At this time, input layers 31a-m+3 to 31
Zero is input to a−m+n. Then, in the neuron unit 31, the weighting coefficient Wji3 and the threshold value θj3 are determined from the input signal r (r=3).
is selected. In this way, the aperture value FT is obtained from the output layer 31c.
, the distance YT to the in-focus position in the case of the passing center spatial frequency ST is determined and divided by Rc in the divider 54, and the distance Y to the in-focus position is detected. Based on this Y value, the motor control circuit 27 drives the motor 17 to move the photographing optical system 2 to the in-focus position.

[0104] In this way, even while the photographing optical system approaches the in-focus position, the weighting coefficients of the neuron units change one after another in accordance with the number of inputs, and the distance Y to the in-focus position is determined moment by moment. When the photographing optical system reaches the position where Y=0, the driving of the motor is stopped and focus adjustment is completed.

Focus adjustment is performed as described above, but since the number of focus signals used for the neural network increases as the photographic optical system approaches the focus position, the value of Y becomes more reliable and more accurate. Focus adjustment is performed with high precision.

Furthermore, since the value of the focus signal and the position of the photographing optical system are normalized, focusing can be adjusted for any aperture value, BPF passing center spatial frequency, and type of subject, and the learning of the neural network is simple. Since only one type of aperture value and BPF passing center spatial frequency are required, learning can be done in a short time, and the number of neurons can be reduced to save memory.

Note that this learning only needs to be performed once for the same type of photographing optical system, and the learning device is not needed during execution. Further, the photographer can use this learning device to perform his own learning, for example, to always perform arbitrary defocusing.

In addition, in the same embodiment, the predetermined steps used for learning were made large to shorten the learning time, but even if the learning time is less than the predetermined step interval at the time of execution, the interpolation effect of the neural network will prevent the learning time from being correct. The in-focus position can be determined. That is, since it may be driven at any step interval, it is possible to drive the photographing optical system at any speed.

Therefore, as a second embodiment, as shown in FIG. 10, the current motor speed is also input to the input layer 31a, and two output layers 31c are used to output the new motor speed. With this configuration, it is possible to learn the ideal motor speed according to the distance to the focus position and the current motor speed, and it is possible to perform focus adjustment not only with high precision but also with high speed. At this time, smooth movement of the photographic optical system may be learned.

Furthermore, in this example, learning started from the position where the photographing optical system always satisfies
There is no need to move to the position of X∞.

[0111]

[Effects of the Invention] As described above, according to the present invention, since a neural network is used to detect the focus position, it is possible to predict the focus position without requiring special conditions, and it is also possible to predict the focus position with high accuracy. Focus position detection becomes possible. Also,
By using the learning device of the present invention, it becomes possible to adjust the focus according to the photographer's intention.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an automatic focusing device according to a first embodiment of the present invention.

2A and 2B are block diagrams each showing a configuration example of the focus signal detector of FIG. 1; FIG.

FIG. 3 is a block diagram showing the configuration of the learning device in FIG. 1;

FIG. 4 is a conceptual diagram of a neural network.

FIG. 5 is a block diagram showing the configuration of the focus position detector of FIG. 1;

FIG. 6 is a diagram showing details of each neuron unit in FIG. 5;

7A and 7B are diagrams each showing a configuration example of the focus signal preprocessing circuit of FIG. 5; FIG.

FIG. 8 is a diagram showing characteristics of a focus signal.

FIG. 9 is a configuration diagram of the photographing optical system position preprocessing circuit of FIG. 5;

FIG. 10 is a block diagram showing the configuration of the focus position detector of FIG. 1 in a second embodiment of the invention.

FIG. 11 is a block diagram showing the configuration of a conventional mountain-climbing automatic focusing device.

FIG. 12 is an operational characteristic diagram of the automatic focusing device of FIG. 11;

FIG. 13 is a block diagram showing the configuration of another conventional autofocus device.

FIG. 14 is an operational characteristic diagram of the automatic focusing device shown in FIG. 13;

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1...Subject, 2...Photographing optical system, 3...Imaging element, 4...Amplifier, 17...Motor, 29a...Photographing optical system information sensor, 3
1... Neural network, 31a... Input layer, 31b
...Intermediate layer, 31c...Output layer, 32...Focus signal detector, 3
3...Focus position detector, 34...Learning device, 35...Distance measurement area designation circuit, 37...Digital integration circuit, 44...Focus signal pre-processing circuit, 45...Photographing optical system position pre-processing circuit, 46...Focus signal storage Circuit, 47...Photographing optical system position memory circuit.

Claims (5)

[Claims] 1. A photographic optical system, an image sensor that captures an image formed by the photographic optical system, and a drive means that changes the relative position of the image sensor and the photographic optical system along an optical axis direction. , an image readout means for reading out the charge accumulated in the image sensor as an image signal, a bandpass filter for passing only a predetermined frequency band from the image signal read by the readout means, and an output signal of the bandpass filter. A focus signal detection means for detecting a focus signal value according to the degree of focus from a focus position information detection means having a neural network, and a focus position information outputted from the focus position information detection means. An automatic focusing device comprising: control means for controlling the driving means accordingly. 2. The focus position information detection means includes a first normalization means for converting a plurality of focus signal values obtained from the focus signal detection means into normalized values, and a first normalization means for converting a plurality of focus signal values obtained from the focus signal detection means into normalized values; a second normalizing means for converting photographing information of the photographing optical system and band information of the bandpass filter into normalized values; and a neural network receiving outputs from the first and second normalizing means. The automatic focusing device according to claim 1, comprising: 3. A photographic optical system, an image sensor that captures an image formed by the photographic optical system, and a drive means that changes the relative position of the image sensor and the photographic optical system along the optical axis direction. , an image readout means for reading out the charge accumulated in the image sensor as an image signal, a bandpass filter for passing only a predetermined frequency band from the image signal read by the readout means, and an output signal of the bandpass filter. a focus signal detection means for detecting a focus signal value according to the degree of focus from the focus position information detection means having a neural network; a learning device for the neural network; An automatic focusing device comprising: a control means for controlling the driving means according to output focus position information. 4. The learning device includes a difference calculator that detects a signal corresponding to a difference between a teacher signal and a learning signal, and a weight that sets a weight between each neuron of the neural network according to the output of the difference calculator. The automatic focusing device according to claim 3, further comprising a setting device. 5. The focus position information detection means includes a first normalization means for converting a plurality of focus signal values obtained from the focus signal detection means into normalized values, and a first normalization means for converting a plurality of focus signal values obtained from the focus signal detection means into normalized values; a second normalizing means for converting photographing information of the photographing optical system and band information of the bandpass filter into normalized values; and a neural network receiving outputs from the first and second normalizing means as input. The automatic focusing device according to claim 3, comprising: an automatic focusing device according to claim 3;