JPH03167536A - Focal point detector for camera - Google Patents

Abstract

PURPOSE:To detect the focal point of main objects in various screens at high speed by allowing a neutral network to learn the distance data vector of numerous model patterns. CONSTITUTION:A detector is provided with the single-layered neutral network 5 consisting of plural neuron units where the distance data vector is inputted through a prescribed synaps combining strength. The neutral network 5 is allowed to learn so as to output position information on the main object on which the focal point of the screen is desired to be adjusted when the distance vector data as the set of each distance data of plural points in the arbitrary screen is inputted, and the focal point is adjusted based on the output. At the time of the learning, the main objects are assumed to be present on the same position when the distance data vectors of the screens are resemble, and the neutral network 5 is allowed to learn each distance vector through the use of the numerous model patterns. Thus, the main object can be determined in a short time in any condition, and the focal point of the main object can be detected at high speed.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a focus detection device for a camera.

[Conventional technology]

In conventional autofocus cameras, a focus frame for distance measurement is located in the center of the screen, and the composition is determined so that the main subject is located there. If the main subject is not in the center or if you want to take a picture with a free composition, you can use focus lock.

In addition, distance measurement is performed not only at one point but at many points on the screen,
There is also a conventional example in which the focus is focused on a point at a predetermined distance, for example, the shortest distance.

Furthermore, we have developed an autofocus control method for cameras that performs pattern folding of many photographs, defines membership functions, and performs approximate inference using fuzzy theory from the relationship between the distance data of three points to determine the position of the main subject. It has been announced.

[Problem to be solved by the invention]

However, in the conventional example where there is only one focus frame, when people, etc. are lined up left and right, the so-called hollow phenomenon cannot be avoided. Further, the focus lock operation is troublesome and is not suitable for quick photography. When focusing on a subject at a predetermined distance from distance data of many points, it is not always possible to take the photograph intended by the photographer.

Furthermore, the disadvantages of applying fuzzy theory are that it is difficult to define the membership function, and the success or failure depends on how it is defined, and because it is an approximate inference, it becomes difficult to process when there is a large amount of distance data to process. The disadvantage is that the time increases rapidly.

This invention was made to address the above-mentioned circumstances,
The purpose is to provide a camera focus detection device with a simple structure that can determine the main subject in a short time under any circumstances and can detect the focus of the main subject at high speed.

[Means and effects for solving the problem] The camera focus detection device according to the present invention uses a single-layer neural network consisting of a plurality of neuron units into which distance data vectors are input via predetermined synaptic connection strengths. Be equipped.

This neural network is trained to output positional information regarding the main subject whose black points on both sides should be aligned when a distance data vector, which is a collection of distance data for multiple points on an arbitrary screen, is input. and adjust the focus based on the output. In addition, when learning,
It is assumed that the main subjects are in similar positions in screens with similar distance data vectors, and a neural ◆ network is made to learn each distance data vector using a large number of model patterns.

FIG. 1 schematically shows a means for determining a main subject using a neural network according to this invention.

The input screen shown on the left is the shooting screen shown in the finder, and the input screen shown in the viewfinder is the area Pi (i-1) arranged in a two-dimensional matrix.
~n (n is a positive integer, shown in Figure 1)),
Distance data Li is obtained for each region Pi. in fact,
The distance of the center point of the region is measured. The neural network shown on the right inputs the distance data vector Li for each area of the input screen and determines the area Px where the main subject is located. Neural used in this invention●
The network is a single-layer network consisting of multiple (25 in this case) neuron units into which distance data vectors are input via predetermined synaptic connection strengths, and each neuron unit represented as a square receives all distance data. A vector is input.

Neural network training consists of the following two steps. First, distance data vectors for a large number of model patterns are input into a neural network, and a predetermined synaptic connection strength is obtained by unsupervised learning. Unsupervised learning is the process of modifying the synaptic connection strength of each neuron unit to selectively respond to a given input vector without any supervised data. automatically categorized above. Next, the output neuron unit of the neural network when the distance data vector of the model pattern is input again to the neural network is associated with the region where the main subject is located in the model pattern. That is,
The area marked in each neuron ● unit in FIG. 1 is the area where the main subject is located.

When photographing, as shown in Figure 1, when the distance data vector Li for each area Pi in the screen to be photographed is input to the neural network trained in this way, any one neuron unit The distance measurement area Px (first
In the figure, P7) can be determined to be the main subject, and if you focus on it, you can always take pictures with the main subject in focus.

The neural network used in this invention will be explained with reference to FIGS. 2 to 4. The model for this neural network is Helsinki University (Finland)
It was proposed by Kohonen and is called a self-organizing network. The learning method is known as self-organizing feature mapping.

Figure 2 shows a model of each neuron unit. The neuron unit is connected to each input value of the input vector via a synaptic connection, compares the input value with each synaptic connection strength, and generates an output according to the degree of matching between the two.

The output y1 of the i-th neuron unit is expressed by the following equation.

yl−f(ΣwljXxkD −(1)4=
1 Here, fO is the output function of the neuron ● unit (sigmoid monotonically increasing function), W1j is the synaptic connection strength of the i-th neuron unit to the j-th input xkj, and xkj is the k-th neuron unit j-th input value of the input vector, n is the number of dimensions of the input vector.

A neural network is constructed by two-dimensionally arranging neuron units, but here, for the sake of simplicity, a one-dimensional model is shown in FIG. 3.

As shown in FIG. 3, this neural network performs self-organizing feature mapping that introduces the effect of feedback coupling of signals from the output of the neuron unit to the input. Here, wlk is the feedback coupling strength from the kth neuron unit to the ith neuron unit.

FIG. 4 shows the degree of influence of such feedback connections on the synaptic connection strength. In other words, due to the effect of feedback connections, the synaptic connection strength exhibits a characteristic change in the shape of a Mexican hat, and when the output value of a particular neuron unit increases, neuron units located in a positional relationship close to it will also be affected by it. The output value of neuron units further outside the neuron unit increases, but the output value of neuron units further outside the neuron unit decreases.

If we simplify the learning process of this neural network (self-organizing feature mapping) so that learning of synaptic connection strength takes into account the effect of feedback connections is performed only in the vicinity of the optimally matched neuron unit, the learning process (self-organizing feature mapping) of this neural network is as follows. It can be done by following these steps.

#1: Initialize W1(0) with a random number and set the number of learning times to 1
-0. Here, Wi(t)−(w11(t).w
12(t). ...wln(t)).

Below, steps #2 and #3 for each input vector Xk
Repeat the process.

#2: Set t-t+1 and find the optimal neuron unit C that satisfies the following equation.

If Xk −we(t)It −i1n {
It Xk −W1(t)If )...(2) Here, Xk − (xkl, xk2, ・−xk
n), w c(t) is the synaptic connection strength of the optimally fitting neuron unit C.

#3: In the case of iεNc(t), W I (t+1)−
W I (t) +α(t) X (Xk-Wl(t
)), and in cases other than i (E N c(t), Wf(
t+1)−Wf(t).

Here, α0〉 is a learning coefficient, and Nc(t) is a set of optimal neuron unit C and its neighboring neuron unit. Usually, both α(t) and Nc(t) are
It is a monotonically decreasing function.

By performing such self-organizing feature mapping, the synaptic connection strengths of neighboring neuron units become similar. Furthermore, since synaptic connection strengths reflecting statistical properties such as similarity and correlation between model patterns are obtained, vector quantization of each model pattern is performed and classification of model patterns becomes possible. Furthermore, the process of determining the position of the main subject in any photographic screen can be considered as the process of determining the optimal fit neuron unit C, but as can be inferred from step #2, this is a fast decision process with a small amount of calculation. Processing is possible.

〔Example〕

Embodiments of the camera focus detection device according to the present invention will be described below with reference to FIGS. 5 to 7. FIG. 5 is a block diagram of an autofocus camera as an example. A plurality of distance measuring sensors 21 (i-1 to n) are provided to detect the distance to each object in a plurality of areas of one shooting screen, and the distance data Li for each area obtained by these sensors determines the main subject. The signal is supplied to the circuit 4 and also to the normalization calculation circuit 3. The normalization calculation circuit 3 normalizes the distance data Li to a real value from 0 to 1 so that it can be input to the neural ● network 5, and obtains the above-mentioned input data vector xkj. The output of the normalization calculation circuit 3 has the synaptic connection strength already acquired through learning, and is input to a single layer neural network 5 consisting of a plurality of neuron units. Here, as the distance data, it is possible to use the distance itself to the subject, or data including distance information, such as the amount of displacement in triangulation method or the amount of lateral shift in phase difference method.

The neural network 5 receives the distance data xkj and calculates the synaptic connection strength w of each neuron unit.
Vector distance N1-Σ(X kj
- w Ij(t) ) 2 is calculated, the minimum value of the distance between the vectors Dc = m1n (Ni l is searched, and the position information Px of the main subject corresponding to the neuron unit that gives the minimum value is output In general, the neural network 5 covers 5 to 2 of the total number of ranging areas.
It consists of approximately 0 times as many neuron units. This positional information of the main subject is input to the main subject determination circuit 4, and the distance data Lx from the ranging sensor 2 corresponding to the main subject is supplied to the focus detection calculation circuit 6 which performs calculations for salt point detection. . The output of the focus detection calculation circuit 6 is supplied to the driver 7, and the driver 7 operates the lens drive system 8 based on this information to move the lens 1 in the optical axis direction and perform focus adjustment.

FIG. 6 shows a specific circuit configuration of the neural network 5. The neural network 5 includes memories 9, 10, 11.12 for storing various data, and various arithmetic circuits 13.14. The memories include a memory 9 that stores the normalized distance data xkj output from the normalization calculation circuit 3, and a synaptic connection strength w1 acquired through learning.
Memory 10 for storing j(t), normalized distance data xk
Vector distance f between j and synaptic connection strength wij(t)
Memory 11 that stores iNi, trained neural
A memory 12 is provided for storing main subject position information indicating the correspondence between neuron units on the network and main subject positions on the m-shadow screen. As the calculation circuits, a circuit 13 that calculates the inter-vector distance Nl from the normalized distance data xkJ and the synaptic connection strength wij(t), and a circuit 14 that detects the minimum value of the inter-vector distance N1 are provided.

Normalized distance data xk output from the normalization calculation circuit 3
j is stored in the memory 9 within the neural network 5. The synaptic connection strength W (1
) is stored in the memory 10 in advance. Memory 9, 10
The data is supplied to the vector-to-vector distance calculation circuit 13, where the vector-to-vector distance calculation is performed independently for each neuron unit, so parallel calculation is possible. Therefore, the inter-vector distance calculation circuit 13 is composed of a plurality of processors, each of which performs processing for each neuron unit, and is capable of high-speed processing. In addition, in a neural network dedicated arithmetic element, that is, a neurochip, each neuron unit is responsible for this arithmetic processing.

The minimum value detection circuit 14 detects the optimum neuron unit C having the minimum value from the obtained inter-vector distance N1. The main subject position determination circuit 4 reads out the main subject position information Px corresponding to this optimal neuron unit from the correspondence stored in the memory 12, and calculates the distance data of the main subject from among the data LL from the distance sensor. Only Lx is supplied to the focus detection calculation circuit 6.

As described above, automatic focus control can be performed on only the main subject among a large number of subjects within the input screen.

In addition, when applying only the learned results, the neural
Although the network is possible with just the configuration described above, initial learning and additional learning when adapting the neural network to the camera user are required. In order to do this, the following learning circuit configuration is required.

FIG. 7 explains the circuit configuration for learning the neural network. This configuration can be realized by simply adding a synaptic connection strength calculation circuit 16 and a main subject position information setting circuit 18 to the circuit shown in FIG.

In this operation, first, the synaptic connection strength wlj of each neuron ● unit of the neural network is initialized with a small random number. Distance data vectors for each region regarding a large number of learning model patterns are stored in the memory 15, and data Px regarding the region where the main subject is located is stored in the memory 17. After determining the optimal neuron unit by detecting the minimum value of the distance between vectors using the distance data in the memory 15 as in the case of FIG. 6, the synaptic connection strength calculation circuit 16 determines the optimal neuron unit. The synaptic connection strength W1 of a group of neuron units in the vicinity of the center is corrected and learned using the following equation. However, iεNc(t). W1(t+1) −Wt(t)+α(t)X (Xk −W1(t)
) ... (3) Here, α(1) is the learning coefficient, which is a real value between 0 and 1, and Nc(t) is the set of neuron units in the vicinity of the optimal neuron unit C. The initial values are not so small compared to the size of the neural network, and α(t), Nc(
t) are both monotonically decreasing functions. t is the same number of learnings, and learning is repeated up to the maximum number of learning times t laX.

After completing the learning for all model patterns, the main subject position information setting circuit 18 determines which neuron unit of the neural network the main subject position of each model shooting screen stored in the memory 17 is associated with. The following relationship is investigated and the result is set in the memory 12 as main subject position information.

yi (=f (ΣwljX x kj) )≧θk
(threshold value in the k-th J.1 input pattern), if the neuron
The main subject position of the k-th model pattern is associated with unit i.

When performing additional learning to adapt a trained neural network to a camera operator, when an arbitrary input pattern is given, the optimal neuron unit is determined by detecting the minimum value of the distance between vectors. , the neuron units that are in the vicinity of that neuron unit and have the same corresponding main object position.
unit! The synaptic connection strength W1 of :1Nc' is corrected and learned by the synaptic connection strength calculation circuit 16 using the following equation.

Wi = Wi + (20
- (4) However, in i6Nc', α0 is a learning coefficient.

This invention is not limited to the embodiments described above, and can be modified in various ways. The learning method explained is unsupervised learning,
When an input pattern is given, the output of the neural network is compared with the training data (the main subject of the input pattern), and if the neural network fires correctly, the synaptic connection strength approaches the input pattern, and if it fires incorrectly, it If so, supervised learning that moves the synaptic connection strength away from the input pattern is also possible. It is also possible to learn by including teaching data as input values. Note that in order to be able to return to the initial learning state before additional learning at any time, it is necessary to store the synaptic connection strength in the initial learning state in memory.

Furthermore, although an autofocus camera was explained as an example,
The present invention may be applied to a manual focus adjustment camera that only auxiliarily displays focus information regarding the determined main subject in a finder or the like.

〔Effect of the invention〕

As explained above, according to the present invention, a neural network is made to learn distance data vectors for a large number of model patterns in processes that were conventionally difficult to formulate and program, such as identifying the main subject. By simply using the present invention, it is possible to provide a camera focus detection device that can quickly detect the focus of a main subject in various screens.

Further speed-up is possible by using parallel processing of neural networks. In addition, by having the neural network learn the camera operator's preferred composition and composition habits and adapting the neural network to the camera operator, the main subject can be determined with higher precision.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing an outline of focus detection of a camera according to the present invention, FIG. 2 is a diagram showing a model of a neuron unit,
Figure 3 is a diagram showing the neural network model,
FIG. 4 is a diagram showing changes in synaptic connection strength due to feedback connections, FIG. 5 is a block diagram of an autofocus camera that is an embodiment of the present invention, FIG. 6 is a diagram showing the circuit configuration of a neural network, and FIG. The figure shows a circuit configuration for learning a neural network. DESCRIPTION OF SYMBOLS 1... Lens, 2... Distance sensor, 3... Normalization calculation circuit, 4... Main subject determination circuit, 5... Neural network, 6... Focus detection calculation circuit, 7
... Driver, 8... Lens drive system.

Claims (1)

[Claims] A device comprising means for measuring distances to a plurality of points within a photographing screen, and a plurality of neuron units, each neuron
The unit includes a single-layer neural network into which a set of distance data for each point output from the measurement means, that is, a distance data vector, is input via a predetermined synaptic connection strength, and a distance data vector of a model pattern. the output of the neural network when the predetermined synaptic connection strength is obtained by modifying the synaptic connection strength so as to cause a selective response, and the distance data vector of the model pattern is inputted into the neural network again; learning means for associating points where the main subject is located in the model pattern; and determining the main subject according to the output of the neural network when a distance data vector of an arbitrary photographic screen is input to the neural network. , and means for detecting the focus accordingly.