JP3217052B2 - Camera focus detection device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a camera focus detection device.

[Conventional technology]

In a conventional autofocus camera, the composition is determined such that a focus frame for distance measurement is located at the center of the screen, and the main subject is located there. When the main subject is not at the center or when the user wants to take a picture with a free composition, the focus lock is used.

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

In addition, an automatic focus control method for cameras that determines the position of the main subject by performing pattern analysis of a large number of photographs, defining membership functions, and performing approximate inference by fuzzy theory from the relationship between the three distance data is also announced. Have been.

[Problems to be solved by the invention]

However, in the conventional example having only one focus frame, when people and the like are arranged side by side, a so-called hollow phenomenon is inevitable. Further, the focus lock operation is troublesome, and is not suitable for quick shooting. When focusing on a subject at a predetermined distance based on distance data of many points, it is not always possible to take a photograph intended by the photographer. Furthermore, the one that applied fuzzy theory is
It is difficult to define a membership function, and its performance depends on how it is defined. Further, since it is an approximate inference, there is a drawback that the processing time increases sharply as the distance data to be processed increases.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and has as its object a simple configuration capable of determining a main subject in a short time under any circumstances and detecting a focus of the main subject at high speed. To provide a camera focus detection device.

[Means and actions for solving the problem]

A focus detection device for a camera according to the present invention includes a distance measurement unit for obtaining a distance data vector that is a set of distance data corresponding to a distance of each of a plurality of regions of a subject image incident via a photographing lens during photographing. , A plurality of neurons connected to each distance data of the input distance data vector via respective synaptic connection strengths.
Units consist of a single layer network,
Each synaptic connection strength has been learned so that the optimal combined neuron unit that responds selectively when the distance data vector corresponding to the model pattern is input,
A neural network for obtaining a value related to the position of the main subject associated with the optimal combined neuron unit obtained for the distance data vector at the time of the imaging by the single-layer network having the plurality of neuron units; Means for adjusting the focus of the camera according to the value of the position of the main subject generated by the neural network.

Further, after the synaptic connection strength is obtained by modifying the synaptic connection strength so as to selectively react to the distance data vector of the model pattern, when the distance data vector of the model pattern is input again to the neural network. And a learning means for associating the output of the neural network with the position of the main subject of the model pattern.

FIG. 1 schematically shows a means for determining a main subject using a neural network according to the present invention. The input screen shown on the left is a photographing screen shown in the viewfinder, and is divided into regions Pi (i = 1 to n (n is a positive integer, 9 in FIG. 1)) arranged in a two-dimensional matrix, and each region Pi Distance data Li is obtained every time. In practice, the distance of the center point of the region is measured. The neural network shown on the right side inputs the distance data vector Li for each region of the input screen and determines the region Px where the main subject is located. The neural network used in the present invention is a single-layer network composed of a plurality (here, 25) of neuron units to which distance data vectors are input via a predetermined synaptic connection strength. A full distance data vector is input to the neuron unit.

Learning a neural network consists of the following two steps. First, distance data vectors in a number of model patterns are input to a neural network,
A predetermined synapse connection strength is obtained by unsupervised learning. Unsupervised learning is the modification of the synaptic connection strength of each neuron unit to selectively respond to a given input vector even without supervised data. Automatically categorized above. Then, 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 described for each neuron unit in FIG. 1 is the area where the main subject is located.

At the time of photographing, as shown in FIG. 1, when a distance data vector Li for each region Pi in the screen to be photographed is input to the neural network trained in this way, one of the neuron units becomes the most important. A strong firing occurs, and the ranging area Px (P7 in FIG. 1) corresponding to the neuron unit can be determined to be the main subject, and by focusing on it, a photograph can be taken in which the main subject is always in focus.

The neural network used in the present invention will be described with reference to FIGS. This neural network model was proposed by Helsinki University of Technology (Finland) and Kohonen,
Called a self-organizing network. The learning method is known as self-organizing feature mapping.

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

The output yi of the ith neuron unit is given by

Where f () is the output function of the neuron unit (usually a sigmoidal monotonically increasing function), wij is the synaptic connection strength for the jth input xkj of the ith neuron unit, and xkj is the kth J is the j-th input value of the input vector, and n is the number of dimensions of the input vector.

The neural network is configured by arranging each neuron unit two-dimensionally. Here, FIG. 3 shows a one-dimensionally modeled one for simplicity of explanation.

As shown in FIG. 3, the neural network performs a self-organizing feature mapping that introduces the effect of feedback coupling of the signal from the output to the input of the neuron unit. Here, wik is the feedback coupling strength from the k-th neuron unit to the i-th neuron unit.

FIG. 4 shows the degree of the influence of such feedback coupling on the synaptic coupling strength. In other words, the synaptic connection strength shows a change in the characteristics of the Mexican hat type shape due to the effect of the feedback connection, and when the output value of a specific neuron unit increases, the neuron units in a positional relationship close to the distance also draw to it. Therefore, the output value of the neuron unit outside the output value decreases.

Simplifying the learning of the synapse connection strength taking into account the effect of this feedback connection so that it is performed only near the optimally matched neuron unit, the learning process (self-organizing feature mapping) of this neural network is as follows. The procedure can be performed as follows.

# 1: Wi (0) is initialized with a random number, and the number of times of learning t = 0. Here, Wi (t) = (wi1 (t), wi2 (t),... W
in (t)).

Hereinafter, the processing of steps # 2 and # 3 is repeated for each input vector Xk.

# 2: Assuming that t = t + 1, an optimal combined neuron unit c satisfying the following equation is obtained.

{Xk−wc (t)} = min {Xk−Wi (t)} (2) where Xk = (xki, xk2,..., Xkn), and wc (t) is c of the optimal combined neuron unit. Is the synaptic connection strength of

# 3: If i∈Nc (t), Wi (t + 1) = Wi (t) +
α (t) × (Xk−Wi (t)), and if i∈Nc (t), Wi (t + 1) = Wi (t).

Here, α (t) is a learning coefficient, and Nc (t) is a set of neuron units near the optimal combined neuron unit c. Normally, both α (t) and Nc (t) are monotonically decreasing functions. It is.

By performing such self-organizing feature mapping, the synaptic connection strengths of neuron units in the vicinity are similar. In addition, since synapse connection strength reflecting statistical properties such as similarity and correlation between model patterns is obtained, vector quantization of each model pattern is performed, and classification of the model patterns becomes possible. Also, the process of determining the position of the main subject on an arbitrary photographing screen can be considered as the process of determining the optimal combined neuron unit c. Processing is possible.

〔Example〕

Hereinafter, an embodiment of a focus detection device for a camera according to the present invention will be described with reference to FIGS. 5 to 7. FIG. 5 is a block diagram of an automatic focusing camera as an embodiment. A plurality of distance measurement sensors 2i (i = 1 to n) for detecting distances to objects in a plurality of regions of one shooting screen are provided, and distance data Li obtained for each region by each region is used to determine a main subject. The signal is supplied to the circuit 4 and to the normalization operation circuit 3. The normalization operation 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 input data vector xkj described above.
The output of the normalization operation circuit 3 has the synapse connection strength already acquired by learning and is input to a single-layer neural network 5 composed of a plurality of neuron units. Here, as the distance data, the distance to the subject itself or data including distance information, for example, a displacement amount by a triangulation method or a lateral displacement amount by a phase difference method can be used.

The neural network 5 receives the distance data xkj, and receives the synaptic connection strength w of each neuron unit.
distance between ij (t) and vector , And the minimum value of the distance between the vectors Dc = min ｛Ni｝
And outputs the position information Px of the main subject corresponding to the neuron unit giving the minimum value. In general, the neural network 5 is composed of about 5 to 20 times as many neuron units as the total number of ranging areas. The position information of the main subject is input to the main subject determination circuit 4 and supplies the distance data Lx from the distance measurement sensor 2 corresponding to the main subject to a focus detection arithmetic circuit 6 that performs arithmetic for focus detection. The output of the focus detection arithmetic circuit 6 is supplied to a driver 7, which operates a lens drive system 8 based on this information to move the lens 1 in the optical axis direction to adjust the focus.

FIG. 6 shows a specific circuit configuration of the neural network 5. The neural network 5 includes memories 9, 10, 11, and 12 for storing various data and various arithmetic circuits 13 and 14.
Consists of The memory 9 stores the normalized distance data xkj output from the normalization operation circuit 3, the memory 10 stores the synapse coupling strength wij (t) obtained by learning, and the synapse coupling with the normalized distance data xkj. A memory 11 for storing a distance Ni between vectors with the intensity wij (t),
Neurons on the trained neural network
A memory 12 is provided for storing main subject position information indicating the correspondence between the unit and the main subject position on the shooting screen. As an arithmetic circuit, a circuit that calculates the inter-vector distance Ni from the normalized distance data xkj and the synaptic connection strength wij (t)
13 and a circuit 14 for detecting the minimum value of the vector distance Ni
Is provided.

Normalized distance data x output from the normalization operation circuit 3
kj is stored in the memory 9 in the neural network 5. Synapse connection strength wij obtained by learning
(T) is stored in the memory 10 in advance. Memory 9, 10
Is supplied to the inter-vector distance calculation circuit 13, where the inter-vector distance is calculated. Can be performed, and the calculation result is stored in the memory 11. Since the calculation of the distance between vectors is independent for each neuron unit, parallel calculation is possible. Therefore, the inter-vector distance calculation circuit 13 is composed of a plurality of processors that perform processing for each neuron unit, and can perform high-speed processing. Further, in the arithmetic element dedicated to the neural network, that is, in the neurochip, each arithmetic processing is performed by each neuron unit.

The minimum value detection circuit 14 detects an optimal combined neuron unit c having the minimum value from the obtained vector distance Ni. The main subject position determination circuit 4 reads main subject position information Px corresponding to the optimal combined neuron unit from the correspondence stored in the memory 12 and, among data Li from all the distance measurement sensors, distance data of the main subject. Only Lx is supplied to the calculation circuit 6 for focus detection.

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

The neural network in the case where only the learned result is applied can be constituted only by the above-described configuration. However, in order to perform the initial learning and the additional learning when the neural network is adapted to the camera operator, the following is required. The learning circuit configuration described in (1) is required.

FIG. 7 illustrates a circuit configuration for learning a neural network. In this configuration, a synapse connection strength calculation circuit 16 and a main subject position information setting circuit are added to the circuit shown in FIG.
It can be realized simply by adding 18.

In the operation, first, the synapse connection strength wij of each neuron unit of the neural network is initialized with a small random number. Stores distance data vectors for each area for a large number of learning model patterns.
In step 15, data Px relating to the area where the main subject is located is stored in the memory 17. After determining the optimal matching neuron unit by detecting the minimum value of the inter-vector distance using the distance data in the memory 15 as in the case of FIG. 6, the synapse connection strength calculation circuit 16 determines the optimal matching neuron unit center. The synapse connection strength Wi of the neighboring neuron unit group is corrected and learned by the following equation. Where i∈Nc
(T).

Wi (t + 1) = Wi (t) + α (t) × (Xk−Wi (t)) (3) where α (t) is a learning coefficient and takes a real value of 0 to 1, and Nc (t) Is a set of neuron units near the optimal combined neuron unit c, the initial value of which is not so small as compared with the size of the neural network, and both α (t) and Nc (t) are monotonically decreasing functions. It is. t is the number of times of learning, and the learning is repeated until the maximum number of times of learning tmax.

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

In the case of (threshold value in the kth input pattern), the main subject position of the kth model pattern is made to correspond to the neuron unit i.

In the case of performing additional learning to adapt to the camera operator in the trained neural network, when an arbitrary input pattern is given, the optimal matching neuron unit is determined by detecting the minimum value of the distance between vectors. The synapse connection strength calculation circuit 16 corrects and learns the synapse connection strength Wi of the neuron unit group Nc 'which is near the neuron unit and has a common corresponding main object position by the following equation.

 Wi = Wi + αo × (Xk−Wi) (4) where i∈Nc ′ and αo is a learning coefficient.

The present invention is not limited to the above-described embodiment, but can be variously modified. The learning method described is unsupervised learning. When an input pattern is given, the output of the neural network is compared with the teacher data (the main subject of the input pattern). If the neural network fires correctly, the synapse is learned. It is also possible to perform supervised learning in which the connection strength is made closer to the input pattern, and if the firing is mistaken, the synapse connection strength is moved away from the input pattern. Also, learning can be performed including teacher data as input values. In order to be able to return to the initial learning state before additional learning at any time, it is necessary to store the synapse connection strength in the initial learning state in a memory.

Further, although the automatic focus camera has been described as an embodiment, the present invention may be applied to a manual focus adjustment camera which merely displays the determined focus information on the main subject in a viewfinder or the like.

〔The invention's effect〕

As described above, according to the present invention, a neural network is made to learn distance data vectors in a large number of model patterns in a process such as identification of a main subject, which was conventionally difficult to formulate and program. Only with this, it is possible to provide a camera focus detection device that can quickly detect the focus of a main subject in various screens. If the parallel processing of the neural network is used, the speed can be further increased. In addition, the neural network learns the desired composition and composition habit of the camera operator, and the neural network is adapted to the camera operator, so that the main subject can be determined with higher accuracy.

[Brief description of the drawings]

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

Claims (2)

(57) [Claims] 1. A distance measuring means for obtaining a distance data vector which is a set of distance data corresponding to a distance of each of a plurality of regions of a subject image incident via a photographing lens at the time of photographing; A plurality of neuron units connected to each distance data of the data vector via each synaptic connection strength are configured as a single layer network, and are selectively selected when the distance data vector corresponding to the model pattern is input. Each synaptic connection strength is learned so as to obtain the optimal combined neuron unit responding to the distance, and the maximum obtained by the single layer network having the plurality of neuron units with respect to the distance data vector at the time of the imaging is obtained. A neural that obtains a value related to the position of the main subject associated with the matching neuron unit A camera focus detection apparatus, comprising: a network; and a unit configured to adjust a focus of the camera according to a value related to a position of a main subject generated by the neural network. 2. After acquiring the synaptic connection strength by modifying the synaptic connection strength so as to make a selective response to the distance data vector of the model pattern, the distance data vector of the model pattern is returned to the neural network. 2. The camera focus detection apparatus according to claim 1, further comprising a learning unit for associating the output of the neural network when input with the main subject position of the model pattern.