JPS59155808A - Automatic focus detecting device - Google Patents

Abstract

PURPOSE:To raise remarkably detecting accuracy and a detecting speed by using an image sensor for covering the whole field of view, and executing a correlating operation by using only a part of a picture data obtained by its sensor. CONSTITUTION:A relay lens 60 for leading a light from a beam splitter 55 to a detecting optical system, a lens 60 for making a pupil, and a pupil splitter 62 placed at a position of a pupil made by the lens 61 are provided on a well- known microscope. Also, a light passing through an image forming lens 63 is made to form an image on an image sensor 65 through a filter 64. Also, a pupil splitter driving circuit 66 and a stage driving circuit 67 are provided, and controlled by a microcomputer 70. Moreover, an image sensor driving circuit 68, and an interface circuit 67 for inputting a picture data from the image sensor 65 to the microcomputer 70 are provided, an automatic focusing operation is executed, a focusing display and an impossibility display are executed by a console 71, and an operation of collelation and a decision of focusing, etc. are all executed by the microcomputer 70.

Description

DETAILED DESCRIPTION OF THE INVENTION This is an automatic focus detection device that converts two images formed through different optical paths into photoelectric output signals by means of an optical converter (a), which is formed by arranging a large number of elements. This invention relates to an automatic focus detection device that performs focus detection by detecting the relative positional relationship between two images based on the optical low output signal.

Conventional automatic focus detection devices of this type include a rangefinder type that applies triangulation and a 'rTL type that divides the light flux passing through the pupil to obtain two images. To detect a match, the correlation is calculated digitally between the two images, and the extreme value of the correlation value is taken as a match, and the relative shift of the two images until the match is taken as the amount of phase difference of one image. Met.

Figure 1 shows an example of this, in which data A and B of two images taken by an image sensor (not shown) are passed through a sample hold circuit, an A-D converter (none of which is shown), etc. The signals are stored in ring-shaped shift registers la and lb, respectively. In this example, the image data has a 128-bit configuration. Both image data A and B are transferred to shift register l.
When stored in a and lb, the absolute value of the difference for each bit is determined by the circuit 2 for determining the absolute value of the difference, and then the sum of these is determined by the adder 3, and the sum of the two images is calculated. It becomes a correlation value. Next, the image data B in the shift register 1b is shifted by one bit by a pulse from the clock OL, and the image data B in the shift register 1b is shifted again to the circuit 2. An adder 3 calculates a correlation value. In this way, a correlation value is determined each time one of the image data is successively moved by the clock OL, and the extreme value of the correlation value is determined by the beak detector 4, and the position where the extreme value is detected is determined as the in-focus position. Become. Further, the number of clocks in the case of an extreme value is obtained by the counter 5, and this number of clocks, that is, the amount of movement of the image data B in the shift register 1b becomes the amount of phase difference between the two images, and from this, the direction and amount of defocus can be known. I can do it.

However, in this conventional device, since the image sensor has to have a certain size, the two images captured on the image sensor are not only shifted in position, but also have different peripheral parts. Result: two shift registers 1. The image data A and B stored in a and lb are not only shifted in position but also different at the ends, and the correlation is calculated while rotating one of the image data in a ring shape. Therefore, accurate image phase difference could not be determined. If this point is explained in detail based on Figure 2,
(a) and (b) show image data A stored in the shift register 1a and image data B stored in the shift register 1b, respectively; when out of focus, both image data A and B shift. Therefore, the beaks P and P' are also shifted, and both image data A and B are different at the ends. Further, (c) shows the image data B in the shift register 1b shifted by α bits. Therefore, it can be seen that when the peaks P and P# match, the phase difference between the two image data A and B is α bit, but among the image data of (C), 0 to
The α bit part is the β~127 bit part in the state (b)! The image data of U, (a) and (c) do not match the entire child. In other words, if the correlation is calculated for all image data from 0 to 127 bits, the correlation value will not necessarily be an extreme value when the image phase difference is zero, that is, when the peaks p and P/7 match. . Therefore, it is difficult to obtain an accurate image phase difference with this device. Also, combined Lf
If you reduce the pitch of the image sensor's elements taking into account the degree of rotation, or reduce the number of elements in the image sensor taking into account the calculation speed, the coverage area of the image sensor becomes narrower and the field of view becomes smaller. There is the trouble of having to bring the object to be focused on in a narrow, limited area.
On the other hand, if the coverage area of the image sensor is widened, the number of elements increases, resulting in a longer calculation time.

In view of the above problems, the present invention uses a photoelectric conversion device that covers the entire field of view, and detects the relative positional relationship using photoelectric output signals of some elements of the photoelectric conversion device, thereby improving detection accuracy. The present invention aims to provide an automatic focus detection device that achieves a significant improvement in detection speed.This will be explained below based on an embodiment shown in FIGS. 3 to 15. shows the principle of the pupil division method used in this example, and in D, (,), 6 is an imaging lens,
7 is a light-shielding plate with an open lower a disposed near the pupil on the front side of the imaging lens 6; 8 is an image plane; when in focus, the image plane 8;
An image Q is formed above, but when out of focus, images correspond to the front and rear bins. The blurred images Q1. Q2 is formed on the image plane 8. (b) shows the case where the opening lower a of the light shielding plate 7 is moved to the opposite side with respect to the optical axis 0. When in focus, an image Q' is formed on the image plane 8, but when out of focus, an image Q' is formed on the image plane 8. Each in front bin.

A blurred image Q+'+Q2' corresponding to the rear bin is formed on the image plane 8. Therefore, for example, the opening lower a of the light plate 7 is (
When moving from position a) to position (b), images Q and q are at the same position when in focus and do not move, but when the front focus is on, the image Ql is empty. It moves to the position 1', and in the case of the rear bin, it moves from Q2 to the position 1g of Q2'. By providing a so-called image sensor on the image plane 8, the state of the image can be measured.

From the above, it becomes possible to distinguish between the front focus and the rear focus, and the amount of out-of-focus at that time can be known from the direction and amount of image movement (ie, phase difference).

FIG. 4 is an overall view of an embodiment in which the above principle is applied to a microscope, where 51 is a light source and 52 is a condenser lens. 53 is a stage on which the specimen is placed for observation. 54 is an objective lens; 55 is a beam splitter that guides light to the focus detection system; 156 is a prism that guides a portion of the light to an eyepiece; and 57 is an eyepiece. 58 is a photographic eyepiece lens, and 59 is a film. Here, the optical system used is no different from a normal microscope. 60 is a relay lens that guides the light from the beam splitter 55 to the detection optical system, 61 is a lens that forms a pupil, and 62 is a lens 6.
This is a pupil divider placed at the position of the pupil created by 1. 63 is an imaging lens, and the light passing through this is filtered 6
4 onto an image sensor 65. 66
is a pupil divider drive circuit, 67 is a stage drive circuit,
Each is controlled by a microcomputer 70.

68 is an image sensor driving circuit; 69 is a microcomputer 70 for transmitting image data from the image sensor 65;
This is an interface circuit that inputs to the Reference numeral 71 denotes a console that performs automatic focusing operations and displays in-focus and impossible indications. The microcomputer 70 performs all correlation calculations, focus determination, etc. For the calculation of the correlation, it is preferable to use a calculation-dedicated LSI that has recently been developed and is now available on the market.

Next, the operation of each part will be explained in detail. Figure 5 shows two specific examples of a pupil splitter that splits the light beam passing through 11 ton to form two images, and (a) shows a glass substrate with a light-shielding part (shaded part). (b) is a sector-shaped device with an opening 10 that is rotated left and right about axis O. As a result, half of the pupils 9 are opened and closed alternately. In the case of (a), D
This method is suitable for a method in which an image sensor is used to capture an image in response to a synchronizing signal from a single-lens divider that is rotated by a C monitor or the like. (b) is suitable for a method in which the pupil divider is moved according to a control device such as a microcomputer and an image is taken by an image sensor in synchronization with the movement of the pupil divider. Thus, by using the pupil divider as described above, (a) in FIG.
, creates the state (b) and its (a).

It becomes possible to obtain image data for each of the states (b) using the image sensor.

Furthermore, since the object or specimen to be focused is generally not always located at the center of the field of view, it is desirable that the image sensor be located as widely as possible and not just at the center of the field of view. However, in order to cover the entire field of view, it is necessary to increase the number of elements in the image sensor. This comes from the fact that the pinch of the element must be made to a certain degree in order to maintain constant focusing accuracy.

This point will be explained below. Figure 6 is a diagram showing the relationship between the amount of phase difference and the amount of defocus between two images.
Now, to simplify the explanation, we will consider a point image. 13 is the optical axis of the optical system, and it is assumed that the point image 11 is formed by the optical system with the number of rear apertures NA'. Assuming that the image sensor 12 is now located at the position of the defocus amount δd, there are two images 11A.

Since lIB is formed with a phase difference of 5p, δd
The relationship between and Sp is Sp δd=−(1) NA′. Now let us consider the focusing accuracy when using a 10x objective lens. lo x NA of the objective lens is 0.40, NA' is 0.04, and from equation (1), δd = 258
p (2) is derived. On the other hand, the depth of focus t of a 10 × objective lens is expressed by ε t = −(3) NA〆 (ε is the circle of least confusion), so ε = 0
．． 05mm (equivalent to a resolution of 20 lines/■), then t=1.25mm (4).

Since focus detection accuracy within this depth of focus is required, δd=0.625 (6
) Therefore, 5p=26 μm (7). In order to obtain this amount of image phase difference with high accuracy, the pitch of the diode play of the image sensor 12 must be approximately 25 μm.

As described above, the pitch of the image sensor 12 is determined based on the requirement for focusing accuracy. In this case, if we use an image sensor with 128 diode arrays on the board,
The range where the image sensor is sharp is 128 x O
, 025 = 3.2 exposure, which is very small compared to the field of view number 21 (field of view is 21 wm in diameter), and the subject to be focused is placed at the position of the image sensor (generally at the center). It becomes necessary to move and focus.

FIG. 7 shows the image sensor used in this embodiment and the processing method used therein. In this example, 512
Image sensor with 7 otodiode arrays
According to this image sensor, 51
2×0.025=12.8 lenghts, and can cover a considerable part of the field of view. However, if the correlation calculation is performed using all bits (diode array), the calculation time will be extremely long and it will be meaningless. So 512 bits are 1
It is divided into five blocks a to d of 28 bits, and correlation calculations are performed using blocks with high contrast.

Here, an example of a contrast calculation method will be described. Generally, the evaluation function for contrast evaluation is 5r(x
) is combined with the sensor's X-bit output, it is known that C-Σ+ r(x)-f (x+1) 1 or 0-Σ(f(X)-f(X+1))" In the case of this embodiment, since it is sufficient to know the relative contrast strength between the contrast marks, which is necessary to accurately know the change in contrast, it is not necessarily necessary to calculate the difference between the bits of the instantaneous shift.For example, C-Σ' Let f(x) − f (x+5) l (Σ' means to calculate every fourth X
), and for example, the calculation for block A is x=64 ==+ r(s4)-r(69) ++・-・・・・・+
+ f (184) - f (188) 1, which means that it is sufficient to add the absolute value of the difference while calculating it 31 times.

The conventional calculation method would require 121 times.

Note that the reason why the absolute value of the difference between the 5-pit, adjacent value, and value is calculated every 4 bits is simply that the absolute value of the difference between the adjacent value is calculated every 4 bits. This is also to improve contrast sensitivity.

In this regard, for example, the image sensor shown in FIG.
When we perform a comparative calculation based on the intensity distribution t of the light above, we find that
In this example, 0=l f (64)-f(69) ++lf(68)
-f. (73) 1=1 13-301+125-60
1 = 52, whereas in the conventional case c = + f (64) - f (65) +++ B68) - f (
69)+=113-141+125-301=6 Therefore, the contrast sensitivity of this embodiment is better than that of the conventional case.

Furthermore, in order to reduce the amount of calculation, if you calculate the difference between the adjacent value of X bits and calculate this every X bits, then X>Y
is preferred. In this embodiment, X=5>4=Y.

Note that the data used for the function f may be either image data A or B.

Thus, blocks a and b are created in the manner described above.

......, the contrast of e is calculated and the one with the best contrast is selected, but as is clear from Figure 7, blocks a and b are 128 to 192 bits apart. 64 so that you don't calculate unnecessary numbers.
~128 bits, 128-192 bits, 192-25
The contrast of 6 bits is calculated respectively, and the contrast of block a is the sum of the contrast of 64 to 128 bits and 128 to 192 bits, and the contrast of block b is 1.
It may also be the sum of the contrasts of 28 to 192 bits and 192 to 256 bits. Furthermore, blocks a, b,
c.

The reason why half of d and e overlap is so that even if there is a part with a significant change in image intensity at the boundary of the block, it is possible to set a block that includes the change. For example, if there is a part where the image intensity changes significantly near the boundary between blocks a and c, that is, around 192 bits, it is not possible to use all the information in block a or block C, but if block b is set, the information Everything is included in block b, which is convenient. Since the calculation for determining the contrast takes a much shorter time than the calculation for the correlation, the calculation time in this embodiment is about 128-bit correlation calculation time + α.

Furthermore, the reason why blocks are not set for each of the 64 bits on both sides is to prevent the situation described in FIG. 2 from occurring when the correlation is calculated by shifting the images.

By performing the processing as described above, even if the subject or specimen to be focused is not necessarily in the center of the field of view, a certain part (block) of the subject is automatically selected and focus detection is performed.

For the above, there is no need to define a fixed block, and the image sensor's many photodiode arrays are arranged so as to cover most of the field of view. The part containing the photodiode array can be selected based on contrast, etc. Marks etc. may still be set manually in the field of view for observation. In this way, even if there is a three-dimensional specimen or dust in the field of view, the user can focus on the subject he or she wants to focus on.

Next, the overall operation will be explained. First, two image data A and B from the image sensor 65 in FIG.
Stored in memory 0.

Then, a block with high contrast among the inner shells of the five blocks is selected, and a correlation is calculated using the image data of that block. Let us proceed by assuming that block a in FIG. 7 is selected.

The correlation calculation is performed on image data A and IIf! corresponding to two images stored in memory. The amount of phase difference between the images is calculated by shifting the image data B one pit at a time and determining how many bits the images need to be shifted to overlap. The correlation formula is, for example, where ABS represents the absolute value and the function t A(x)r
fB(x) represents the value of the image data A and BOX bit, respectively. Then, − set of functions f A, +
When δ is changed for f B , δ when R(δ) is minimized, that is, δl, is taken as the phase difference. Further, in this example, -64≦δ≦64. Since the range of δ can be narrowed near the focal point, calculation time can be shortened.

Since the actual value of δ is only taken for each bit of the image sensor, in order to test with even higher accuracy, the value of the correlation is approximated by curve fitting, etc., and the accuracy is less than 1 pit. Find the phase difference of the images (Figure 9). Alternatively, it can also be determined by performing quadratic curve approximation using the one point δ where the value (δ) is minimum, that is, δ', and three points '+p+Q before and after it (Figure 10).

As a result of the above, since many parts of the field of view can be sharpened and focus accuracy can be maintained, the calculation time can be hardly increased.

In the above example, δ is in the range of -64 to 64, and the defocus amount in this range is 0f325'18←4 from equation (6).
vehicle, and on the objective side 40/10" = 0.4 = 400
Since it is μm, the defocus amount is ±400 μm.

If an attempt is made to include a wider range of defocus amounts within the detection range, it may be possible to increase the range of δ, but this increases the amount of calculations and is not preferable.

Also, when the amount of defocus is large like that, there is no point in performing highly accurate calculations like the one I did above.

Therefore, in this embodiment, the forward calculation of contrast and correlation is performed using data of every few bits of the data taken in as image data. Specifically, if data is used every 5 bits, data f(0
), f(1), f(2),...f (510)
, f (511) of f(0), f(5), f
(10), . . . f (505), r (slo) can be considered as the data to be used. In fact, when calculating
It simply uses bit-by-bit data. For example (8)
The formula is 64 If (96+4n)-f (96+4n+δ
)+ (9)n:OAB. In this case, since the defocus amount is large and the image is blurred and contains only low frequency components, no block division is performed. In this case, the amount of change in δ is also calculated every five, so even if the range of δ is, for example, −200≦δ≦200, the number of correlation calculations can be as small as 81 times. The detection range is ±1.25■.

As described above, by using image data every few bits, the detection range can be expanded without increasing the amount of calculation. When the amount of defocus is large like this, by calculating every few bits to get close to the focus position, and then performing calculations that take into account the above-mentioned coverage, the focus range will be wide and the focus coverage will be good. Automatic focus detection can be performed.

Furthermore, if the amount of defocus is large, move the stage while calculating only the contrast (which requires less calculation) from the image data, and then perform focusing by correlation after the contrast reaches a constant value. good. However, since there is a risk that the correlation calculation will calculate an incorrect image phase if the contrast is not higher than a certain level, it may be used to determine whether or not to perform the calculation if the contrast is lower than that.

In addition, in the case of a device such as a camera that moves an objective optical system to perform focusing, it is not necessary to drive the optical system.

In the case of the above embodiment, the light flux passing through the pupil is divided to obtain two images, so the image data A
and B may be different in amount.

This is especially likely to be the case if the focusing system is attached. If there is no lli divider at the pupil position, the 11th
As shown in the figure, there is unevenness in the amount of light between image data A and B. FIG. 12 is a diagram schematically illustrating unevenness in light amount. (a)
In the case of , the pupil and pupil divider match, so each image height is ``+'
For +j, the amount of light passing through the pupil is all equal to a. (b
), since the pupil and the pupil divider do not match, the amount of light passing through the pupil for the valley image height hr'*J is s, a, and c, respectively, and is non-uniform, resulting in uneven light amount.

If there is a difference in the amount of light or unevenness in the amount of light as described above, the image data A
The similarity between the two images B and B will deteriorate, and the accuracy of the correlation processing result will drop significantly. Therefore, correction is necessary. One example of correction is a method often used to remove fixed pattern noise from image sensors, in which image data is obtained by exposing an image sensor to uniform light in advance.

Due to the uniformity of the incident light, the image data is just fixed-noise noise, so if you create a correction coefficient using its reciprocal, then by multiplying the correction coefficient by the image data, you can obtain a fixed-noise noise. It is possible to remove the influence of noise. In the case of this embodiment, when image data A and B are obtained with uniform light passing through the focusing optical system, image data A and B become data with uneven light intensity as shown in Fig. 11 due to eccentricity, etc. ing. Therefore, by creating a correction coefficient using its reciprocal or the like and performing similar processing, the influence of uneven light amount can be removed. Even if there is a difference in light intensity, the same effect can be obtained by performing the same processing through the optical system. Incidentally, it also serves to eliminate fixed chime turn noise of the image sensor. As a specific method for exposing with uniform light, Stage 5
A simple method is to input image data without placing a sample on the sample.

As described above, for the above correction, it is necessary to input data using uniform light once. Although this is not a troublesome operation, it still requires a lot of data input for correction (if necessary, correction can also be done by calculation. Figure 13 is a diagram explaining this, with the The y-axis is in the direction of the intensity of the image data.The values of the image data A and B can be thought of as straight lines with a constant slope, as can be seen from the explanation of FIG.

Each is set as 1.1. If the slope B of image data ZA is β, then il! II image data tAO formula is I
If A is the average light amount of tA, then y = βAX + engineering A. Here, the slope β changes depending on the amount of light, but
It can be found from The gong is determined by the characteristics of its optical system, so it is best to measure it in advance. If the average of the light amounts of the image data A and B is ``att'', then the correction coefficient α is equal to .alpha., which makes it possible to correct the difference in light amount and the unevenness of light amount.

As described above, by using uniform light or by multiplying by 1, it is possible to correct and eliminate the effects of eccentricity of the optical system or eccentricity of the pupil, or the effects of misalignment between the pupil and the pupil divider. I can do it. As a result, focusing accuracy is improved and the detection range is expanded. Furthermore, the focus detection section can also be used as an attachment. The most significant effect is that objective lenses with different magnifications can be used for each of the 11 different positions.

Furthermore, when many types of objective lenses are used, such as in a microscope, the IRn position differs depending on each objective lens, so it is difficult to correctly install the pupil divider at the pupil position for all the objective lenses. One possible solution to this problem is to provide a plurality of pupil dividers. By providing a plurality of pupil dividers at the pupil positions of respective objective lenses, the pupils and the pupil dividers are made to coincide. It goes without saying that when one +Ii divider is used, the other pupil dividers are configured so as not to disturb the pupil. For example, the 14th
As shown in the figure, it is also possible to use the same two pupil dividers as in FIG. 5 connected together.

In FIG. 4, the filter 64 is an infrared cut filter or a band filter, and the spectral sensitivity of the image sensor and the amount of light 51, etc., is different from the relative sensitivity by 1 spectral sensitivity, so the focus shift is caused by this. It serves to prevent the phenomenon of

When performing control and arithmetic processing of the automatic focusing device as described above, a method using a microcomputer and an arithmetic processing unit is easiest and cheapest in design.

Regarding this, the explanation will be supplemented with the flowchart shown in FIG. 15. This shows the most basic case. When focusing is started, it is first checked to see if the microscope is in a state suitable for focusing, and the type and magnification of the objective are determined. This is because in the case of light intensity unevenness correction, the noclameter varies depending on the type of objective and the magnification, and the conversion coefficient for converting the amount of image phase difference into the amount of stage movement also differs depending on the magnification (see equation (1)). Next, the data of fB and fB are obtained from the image sensor and stored in the memory. after that,
Correct the uneven light intensity and store it in memory again. Since the focus may be largely off at the start of focusing, the approximate focus position is determined by correlation calculation every 5 bits ((9)
(see formula). Then, the image phase difference amount determined based on the correlation is converted into a stage movement amount, and the stage is moved. So again f
Obtain data of A and fB and perform correction. Next, blocks are determined by contrast evaluation. Since the reliability of the results due to correlation is a little low unless the control value exceeds a certain value,
Perform 5-bit correlation again and move the stage closer to the focus position. If the contrast does not increase even after repeating several times, the contrast of the sample is too low, and an "impossible" indication is displayed. If the contrast is above a certain level, the correlation is calculated using the determined block, and the stage is moved to focus. To confirm focus, fA and fB are obtained again and the correlation is calculated. Here, if the image phase difference amount is within the depth of focus, the image is in focus and the stage is not moved. If it is outside the depth of focus, repeat the same operation again.

The above is an explanation of the most basic operation, and the actual program and RAM take into account cases where there is no sample, fail-say 7, etc. when the machine breaks down.

Further, the image phase difference view may be used for determination when changing from 5-bit correlation to 1-bit correlation. In the case of the previous example -200≦δ≦
200 ranges are calculated every 5th range, but the correlation R(δ
) takes the minimum value when δ is 1200≦δ'≦200, the stage is moved by that amount and then the process moves to 1-bit correlation. In this case, it is better to set the criterion to be smaller than the calculated range of δ, such as -180≦δ'≦180. This is because if the defocus is large, noise or the like may cause the δ to take the minimum value to be determined incorrectly.

FIG. 16 shows a second embodiment of a control/arithmetic circuit in which components other than the central processing unit are constructed of hardware. To explain this, first, the pupil divider drive Io circuit 66 operates in response to a focus start signal from the console 71 to obtain image data A by the image sensor 65. The image sensor 65 starts imaging by the sensor drive circuit 68 in synchronization with the pupil splitter 62. At this time, in the case of an accumulation type image sensor (generally, solid-state image sensing devices are of this type), a color readout is performed in order to erase the accumulated signals before evacuation. Image data continuously read out from the image sensor 65 is stored in a sample and hold circuit 31. A, /D converter 32. The first memory 3 passes through the switch circuit 33
4 is stored. Then, 1. is written in advance in a memory (not shown). The image data A is corrected using the corrected correction coefficient data and stored in the first memory 34 again. The correction coefficient data is obtained by multiplying the reciprocal of the above-mentioned image data captured by an image sensor with uniform light incident thereon by the average value of the image data A and B at that time. Consider the case of a 512-bit image sensor. Image data A
, B has a total of 10 bits from 0 to 1023 when both are installed.
There will be 24 pieces. Image data A is 0 to 511 bits, and image data B is 1512 to 1023 bits. If the value of the n-th bit of image data obtained with uniform light without a sample is Xn, then the correction coefficient kn of n bits is 023 ΣX.

Next, when the image data A is stored in the first memory 34-, the pupil divider 62 becomes in a state to take the image data B and performs the same process as the image data A.

It is corrected and stored in the second memory 35. 1st memory 3
The data stored in 4 is sent to the contrast discriminator 36 for each block shown in FIG. 7, and the block to be used is determined based on the height of contrast. If block b has the highest contrast, addressing circuit 37
128 is given to . The address shift circuit 38 contains an initial value -32, and when the address designation circuit 37 designates 128 in the first memory 34, the address shift circuit 38 designates 96 in the second memory 35. Then, the image data fA (128) and fB (96) are input to the subtracter 39 and passed through the absolute value circuit 40 to l fA (128) - fB.
(96) An operation of 1 is performed. And adder 41
The data is then stored in the memory 42. When the above is completed, the address designating circuit 37 designates 129 in the memory 34, the address shift circuit 38 designates 97 in the memory 35, and the same calculation is performed thereafter as +fA(t29)-rB(97)+
is the previous data + fA (12g) by the adder 41
−rB(96) + and stored in the memory 42. Thereafter, the process is repeated up to address 255 in the memory 34, and the phase is completed. When this calculation is completed, the value of the address shift circuit 38 becomes -31, and the following is calculated. And this is the address shift circuit 38
This continues until the value of becomes 31, and the entire correlation calculation is performed. Next, the central processing unit 43 processes the shift memory 4.
2, R(δ) is compared, and δ that minimizes R(δ) is found and used as the amount of phase difference of the image. Accordingly, the stage drive circuit 67 is driven to adjust the focus.

If the defocus amount is insufficient with the phase difference amount of 32≦δ≦31, the address designation circuit 37 specifies 128i of the first memory 34 and the address shift circuit 3
The initial value of 8 is -64. Then, each time the correlation calculation is completed, it increases by 2, and R (-64) = I information (12
8) -t 8 (128-64) + Toge A, (130) -
fB(130-64)l+−・・・・・−−・・・・+
l fA(382)-fB(3B2-64) 1R(-
62)-...-... n(60)=... The following calculation is performed. This corresponds to using every other bit of the image data, and the detection range of the image phase difference is doubled with the same amount of calculation. However, the focusing accuracy becomes 1/2.

As described above, the image phase difference is calculated and the stage drive circuit 67 is driven to adjust the focus, but the above operation may be repeated several times to accurately adjust the focus. Furthermore, console 7
In step 1, focusing is started and focused display is performed.

As mentioned above, the automatic focus detection device according to the present invention uses an image sensor that covers the entire field of view, and performs correlation calculations using only a part of the image data obtained by the image sensor, so the detection accuracy can be improved. and the detection speed is greatly improved.

It goes without saying that the present invention can also be applied to an automatic focus detection device using two image sensors.

[Brief explanation of drawings]

Figure 1 is a diagram showing the main circuit of a conventional automatic focus detection device.
FIG. 2 is a diagram showing an example of a part of the image stored in the conventional shift register, and FIG. 3 is a diagram showing the principle of the pupil division method used in an embodiment of the automatic focus detection device according to the present invention. , FIG. 4 is an overall view of the above embodiment, FIG. 5 is a front view showing a specific example of a one-fail divider used in the above embodiment, and FIG. 6 is a diagram showing the phase difference and defocus amount of two images. 7 is a diagram showing the image sensor of the above embodiment and its processing method, FIG. 8 is a diagram showing an example of the light intensity distribution on the image sensor, and FIGS. FIG. 10 is a diagram showing a highly accurate phase difference calculation method near the in-focus point, FIG. 11 is a diagram showing unevenness in the amount of light of image data in the above embodiment, FIG. 12 is a diagram schematically explaining the unevenness in light amount, and FIG. 1
FIG. 3 is a diagram showing a method for correcting light amount unevenness, FIG. 14 is a perspective view of another pupil divider, FIG. 15 is a flowchart showing a computer control and arithmetic processing method of the above embodiment, and FIG. 16 is a diagram showing another method FIG. 3 is a diagram showing a control/arithmetic circuit according to an embodiment. 51... Light source, 52-. Condenser lens, 53.
...Stage, 54...Objective lens, 55--Beam splitter-156...Prism, 57--Eyepiece lens, 58--Photography eyepiece lens, 59--Film, 60--IJL /-V7's, 61-...lens,
62...l ramming device, 6 discharge...imaging lens, 64...
...Filter, 65...Image sensor-166.
... Pupil divider I11 movement circuit, 67... Stage drive circuit, 68... Image sensor drive circuit, 69...
- Interface circuit, 70... microcomputer, 71...-z y 7- /l/. Agent Yasushi Shinohara Jishi 1", (G Boiled Figure 11 Figure 2 Figure 3'3 Figure 3'9 Figure 10 R (X)! Figure 11

Claims (1)

[Claims] Two images formed through different optical paths are individually converted into photoelectric output signals by a photoelectric conversion device comprising a large number of elements arranged, and the relative positional relationship between the two images is detected based on the photoelectric output signals. In an automatic focus detection device that performs in-focus detection, a photoelectric conversion device that covers the entire field of view is used, and a photoelectric output signal of some elements of the photoelectric conversion device is used. An automatic focus detection device characterized by detecting a relative positional relationship.