JPH09134433A - Image recognition system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a system capable of performing an image recognition with high recognition accuracy and in real time and low in cost by providing a single density detection unit for correcting the fluctuation of the illumination at the time of inputting image information between an input layer and an intermediate layer. SOLUTION: A multilayered type neural network composed by hierarchizing an input layer 22, an intermediate layer 23 and an output layer 24 is used as an image recognition means 20. A single density detection unit 25 for correcting the fluctuation of the illumination at the time of inputting image information is provided between the input layer 22 and the intermediate layer 23. The single density detection unit 25 is composed by being coupled with each unit of the intermediate layer 23 by the weight which is preliminarily learned for each unit composing the intermediate layer 23 and by being coupled with each unit of the input layer 22 by the weight which is preliminarily learned in each unit composing the input layer 22. Thus, the extraction accuracy of the feature necessary for the discrimination of an image can be secured, regardless of the fluctuation of the illumination.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The image recognition apparatus of the present invention uses a multi-layered neural network in which an input layer, an intermediate layer, and an output layer are layered to perform image recognition for input image information. The present invention relates to an apparatus, and more particularly, to an image recognition apparatus including a density detection unit for correcting illumination fluctuation.

[0002]

2. Description of the Related Art As a conventional image recognition device 9 of this type, there is one shown in FIG. 4, for example. In order to reduce the number of pixels for which image processing is performed after the optical image information from the optically illuminated target object 1 is converted into an electric signal as image information by the image input unit 2b provided in the input means 2. Mosaic processing of about 12 × 12 pixels has been performed.

Further, in order to ensure the extraction accuracy of the features required for identification, the preprocessing unit 2c installed in the input means 2
Some pre-processing was previously performed on the optically input original image information. As preprocessing, specifically, so-called gradation normalization of about 256 gradations in black and white has been executed in advance.

The preprocessed image information has been input to the image recognition means (that is, a neural network) 3 via the delay line 2d provided in the input means 2. Image recognition means (that is, neural network) 3
In, each unit forming the input layer 3a is connected to each of all the units forming the intermediate layer 3b by a weight learned in advance for all units forming the intermediate layer 3b, and each unit forming the intermediate layer 3b is , Output layer 3c
Were combined with each of all the units by pre-learned weights.

The image recognizing means (that is, a neural network) 3 having such a configuration inputs according to a predetermined calculation and weights between layers which are taught and learned in advance by using a learning rule such as backpropagation. The recognition of the mosaicked image information was performed.

[0006]

However, in the conventional image recognition means 3 as described above, in order to secure the extraction accuracy of the features necessary for the identification, the original image information is not subjected to the gradation normalization or the like. Since it is necessary to execute the processing in advance, it takes a long processing time. In order to perform highly accurate image identification in real time,
There is a problem that a high-speed image recognition means and a high-speed and high-accuracy imaging device are required.

Further, the method of extracting a feature from grayscale image information such as image information and mosaiced image information that has been preprocessed such as gradation processing strongly depends on the illumination condition of the target object, and is therefore dependent on the illumination variation. There was a problem that it was weak. The present invention has been made in view of such a conventional problem, and uses video signal information as input image information to be input to the input layer instead of image information that has been preprocessed such as gradation processing. In addition to adopting the method, a single density detection unit for correcting the fluctuation of illumination when inputting an image is provided between the input layer and the intermediate layer, and the extraction accuracy of the features required for image identification is adjusted to the illumination fluctuation. It is an object of the present invention to provide an image recognition device that can be secured regardless of the above, and that can perform image recognition with high recognition accuracy and in real time and that has a low device cost.

[0008]

Means for achieving the object are as follows. [1] Item In the image recognition apparatus 10 that recognizes input image information by using a multilayer neural network in which an input layer 22, an intermediate layer 23, and an output layer 24 are layered as the image recognition means 20. The image recognition device 10 is characterized in that a single density detection unit 25 is provided between the input layer 22 and the intermediate layer 23 to correct a variation in illumination when inputting image information.

[2] Item The single concentration detecting unit 25 is connected to each unit constituting the intermediate layer 23 by a weight learned in advance for each unit, and also constitutes the input layer 22. The image recognition device 10 according to the item [1], which is formed by combining each unit with a weight learned in advance for each unit.

[3] Item Each unit constituting the input layer 22 is composed of the intermediate layer 2
3 is combined with each of all the units by the weights learned in advance for all the units forming 3, and each unit forming the intermediate layer 23 is learned in advance about all the units forming the output layer 24. The image recognition device 10 according to item [1] or [2], wherein the image recognition device 10 is combined with each of the units by different weights.

[4] Item [1] to [1], wherein the input / output function of the single concentration detecting unit 25 is a sigmoid function 25A.
The image recognition device 10 according to the item [3].

[5] Item Weight between the single concentration detecting unit 25 and each unit forming the intermediate layer 23, or between the single concentration detecting unit 25 and each unit forming the input layer 22. The image recognition device 10 according to the items [1] to [4], wherein the weights are learned in advance so that the input value to the sigmoid function 25A falls within a predetermined range.

[6] Item The intermediate layer 23 can correct the illumination variation by changing its own amplification factor in accordance with a change in the output from the single concentration detection unit 25. The image recognition device 1 according to items [1] to [5].
0.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a functional block diagram showing an image recognition device 10 according to an embodiment of the present invention. FIG. 2 shows a single concentration detection unit 2 according to the embodiment of the present invention.
5 is a graph showing characteristics of an input / output function (sigmoid function 25A) of No. 5; FIG. 3 is a graph showing the relationship between the output value and the image gradation value of the single density detection unit 25 according to the embodiment of the present invention.

First, the configuration of the image recognition apparatus 10 according to the embodiment of the invention will be described. The image recognition device 10 according to the embodiment of the present invention includes an input layer 22, an intermediate layer 23, and an output layer 2.
In a device that recognizes input image information by using a multilayer neural network in which 4 is hierarchized as the image recognition unit 20, as shown in FIG. An input means 30 for reading and converting the input image information into an electrical input image, and an image recognition means 20 capable of performing recognition of the image information on the input image information and identifying the target object 11 are provided. Consists of

The target object 11 is illuminated by an optical illumination means (not shown). The optical lens system 31 mounted on the input means 30 is used for illuminating the target object 11
This is for focusing the optical image information on the image input unit 34 of the next stage.

An image input section 34 provided downstream of the optical lens system 31 converts optical image information of the target object 11 into electrical input image information. Specifically, 12
A photoelectric conversion device such as a CCD area sensor having × 12 = 144 pixels is used.

A signal amplification processing unit (specifically, a video amplifier) 32 provided in a stage subsequent to the image input unit 34 executes a video signal amplification process for electrically amplifying the image information generated by the image input unit 34. Then, the video signal information is generated. The image input unit 34 according to the embodiment of the present invention is specifically an electronic circuit for image input such as a video amplifier and computer hardware for executing image processing of input image information. Can be realized.

The delay line 33 converts the serially transmitted video signal information into a parallel signal and inputs it to the input layer 22 of the image recognition means (multilayer neural network) 20.
Is transmitted to each unit. The delay line 33 according to the embodiment of the present invention can be specifically realized by a semiconductor IC.

Next, the configuration of the image recognition means (multilayer neural network) 20 according to the embodiment of the invention will be described. The image recognizing means (multilayer neural network) 20 according to the embodiment of the present invention uses a predetermined calculation (that is, a neuro calculation) between the weights between the layers that are taught and learned in advance by using a learning rule such as backpropagation. ) And
As shown in FIG. 1, the input image information recognition unit includes an input layer 22, a single density detection unit 25, an intermediate layer 23, and an output layer 24.

The image recognition means (multilayer neural network) 20 according to the embodiment of the present invention can be specifically realized by a program on a computer. Each unit forming the input layer 22 is combined with each of all the units of the intermediate layer 23 by weights learned in advance for all the units forming the intermediate layer 23.

The number of units forming the input layer 22 is the same as the number of pixels of the image input unit 34 (that is, 12 × 12 = 144). In the embodiment of the present invention, 14
It is set to 4 units. The input image information input to the input layer 22 has a predetermined number of pixels (specifically, 12 × 12).
= 144).

Each unit forming the intermediate layer 23 is connected to each of all the units forming the output layer 24 by a weight learned in advance for all the units forming the output layer 24. The number of units forming the intermediate layer 23 can be variable, but is set to 35 units in the embodiment of the present invention.

Each unit forming the output layer 24 is combined with each of all the units forming the intermediate layer 23 according to the weight learned in advance. The number of units forming the output layer 24 is equal to the number of types of the target object 11 to be identified. In the embodiment of the present invention, about 12 units are set so that it can be applied to the determination of coins and the like.

A single density detecting unit 25 for correcting the fluctuation of the illumination when the image information is input is provided between the input layer 22 and the intermediate layer 23. The single concentration detection unit 25 is coupled to each unit of the intermediate layer 23 by a weight previously learned for each unit of the intermediate layer 23, and is also pre-learned for each unit of the input layer 22. Each unit of the input layer 22 is connected to each unit by the weight.

Specifically, a single concentration detection unit 25
Is coupled to each unit of the intermediate layer 23 by the weight learned in advance for each unit forming the intermediate layer 23, and the input layer 22 is connected by each weight learned in advance for each unit forming the input layer 22. Each unit is connected to each unit.

As an input / output function of the single concentration detecting unit 25, a sigmoid function 25A as shown in FIG.
Is used. Next, the operation of the embodiment of the invention will be described. First, the operation of the image recognition apparatus 10 according to the embodiment of the invention will be described.

Image recognition apparatus 10 according to the embodiment of the present invention
Can use the multi-layered neural network in which the input layer 22, the intermediate layer 23, and the output layer 24 are hierarchized as the image recognition means 20 to execute the recognition of the input image information. The multi-layered neural network can perform information processing by imitating the mechanism of a neural network of higher animals. As a feature of the application using the multi-layered neural network 20, the learned network can obtain the output without complicated calculation.

The recognition of image information is to extract a feature required for identification from given input image information and assign an object to a given category from the feature. Each category corresponds to each unit of the output layer 24. Normally, the same number of units as the number of categories are set in the output layer 24. The category assignment result can be known by detecting which unit in the output layer 24 is excited.

Further, the image recognition apparatus 10 according to the embodiment of the present invention uses the video signal information as the input image information input to the input layer in place of the image information which has been subjected to the preprocessing such as gradation processing. , It is possible to secure the extraction accuracy of the features necessary for image identification regardless of the illumination fluctuation.

In the embodiment of the present invention, the number of pixels is 12 ×.
12 = 144. Therefore, the number of divisions is the target object 1
Although depending on the complexity (specifically, the fineness) of 1, the minimum number of divisions is used in order to simplify the calculation and make the image recognition apparatus 10 to be put into practical use inexpensive and simple. ing.

As a result, it is possible to realize an image recognizing device which has a high recognizing accuracy and can identify an image in real time and which has a low device cost. Further, by providing a single density detection unit 25 for correcting the fluctuation of the illumination at the time of inputting the image information between the input layer 22 and the intermediate layer 23, the extraction accuracy of the features necessary for the image identification is improved. It is possible to realize the image recognition device 10 that can be secured regardless of the above.

Next, the operation of the input means 30 according to the embodiment of the invention will be described. The target object 11 is illuminated by an optical illumination means (not shown). Input means 3
The optical lens system 31 installed at 0 can focus the optical image information of the illuminated target object 11 on the image input unit 34 of the next stage.

An image input unit (specifically, a CCD area sensor having 12 × 12 = 144 pixels) 34 provided in the subsequent stage of the optical lens system 31 receives the optical image information of the target object 11 as an electrical input image. It can be converted into information.
The signal amplification processing unit 32 installed in the latter stage of the image input unit 34
Can generate video signal information by executing a video signal amplification process for electrically amplifying the image information generated by the image input unit 34.

The delay line 33 converts the serially transmitted video signal information into a parallel signal and inputs it to the input layer 22 of the image recognition means (multilayer neural network) 20.
Can be transmitted to each unit. Next, the operation of the image recognition means (multilayer neural network) 20 according to the embodiment of the invention will be described.

The image recognition means (multilayer neural network) 20 according to the embodiment of the present invention uses the weights between the layers, which are preliminarily taught and learned by using a learning rule such as backpropagation, and a predetermined calculation (that is, a rule). , Neuro operation) and recognition of the input image information can be performed.

The determination of the identification is performed by checking the output value of the output layer 24 when the image information is presented and checking whether or not the corresponding unit shows the maximum value. In addition, as a criterion, the conventional technique is used, and three criteria are set: correct answer, error, and not applicable. Not applicable means the maximum output is max1 and the second output is max.
2, when max1 <0.4 or (max2-m
This is the case where ax1) / max1 <0.2.

The image recognition means (multilayer neural network) 20 of the embodiment of the present invention is specifically realized by a program on a computer. The weight between the single concentration detection unit 25 and each unit constituting the intermediate layer 23, or the single concentration detection unit 25
As for the weight between each unit constituting the input layer 22 and the input layer 22, the input value to the sigmoid function 25A is within a predetermined range (that is, FIG.
Is pre-learned so that it fits within the non-linear portion of the broken line).

In the sigmoid function 25A, the input of the sigmoid function 25A is made negative as shown in FIG.
The non-linear output part is used. Further, as shown in FIG. 3, the output of the single density detection unit 25 decreases exponentially as the overall gradation increases.

The intermediate layer 23 is a single concentration detection unit 2
Illumination fluctuation can be corrected by changing the amplification factor of itself according to the change of the output from 5. next,
The learning operation of the embodiment of the invention will be described. Output layer 24
Is provided with units for the number of target objects 11 to be identified,
Correspond to each. Back-propagation learning is performed by using the individual identification signal corresponding to the video signal information as a teacher, giving 0.9 to the target output unit and 0.1 to the other units, and repeating this.

The learning process is shown below. [Process 1]: Teacher signal Tk corresponding to presentation pattern
And the output Ok of the output layer 24, the error δk with respect to the weight connected to the output layer 24 and the threshold of the intermediate layer 23 is obtained.

[Process 2]: As the initial values of the weight and the threshold value, a random number value from -0.3 to 0.3 is given. Also, the coefficients α and β during the back propagation learning
Is set to 1.0.

[Process 3]: For all target objects 11,
When the mean square error between the output value and the correct value in the output unit of the output layer 24 is within a certain value (that is, when the learning has converged), the learning ends.

When the weights and thresholds are modified, an inertial term may be given in order to reduce vibration and accelerate learning convergence. The multi-layered neural network 20 learned according to such an algorithm corrects the weight and threshold value of the network, and constructs the feature extraction capability required for identification in the network.

Through the above process, the learned multilayer neural network 20 is
You have the ability to associate what the object is with the category. In such an image recognition means (multilayer neural network) 20, a feature required for identification is extracted in the intermediate layer 23, and the output layer 24 determines an output from the feature.

As described above, according to the embodiment of the present invention, the single density detecting unit 25 for correcting the fluctuation of the illumination when the image information is input is provided with the input layer 22 and the intermediate layer 23.
By providing the image recognition apparatus 10 between and, it is possible to realize the image recognition apparatus 10 at a low apparatus cost, which is capable of ensuring the extraction accuracy of the features required for the image recognition regardless of the illumination variation and capable of performing the image recognition in real time. .

More specifically, in a field where real-time identification processing such as product inspection in a factory is required, image information can be obtained in real time by using an image recognition device 10 that does not require much device cost such as a high-speed personal computer. Will be able to identify.

[0048]

According to the present invention, by providing a single density detection unit for correcting the fluctuation of illumination when inputting image information between the input layer and the intermediate layer, the characteristics required for image identification are provided. The extraction accuracy can be secured regardless of lighting fluctuations,
Image recognition is possible with higher recognition accuracy and in real time.
An image recognition device with low device cost can be realized.

[Brief description of the drawings]

FIG. 1 is a functional block diagram showing an image recognition device according to an embodiment of the present invention.

FIG. 2 is a graph showing characteristics of an input / output function (sigmoid function) of a single concentration detection unit according to the embodiment of the present invention.

FIG. 3 is a graph showing the relationship between the output value and the image gradation value of a single density detection unit according to the embodiment of the present invention.

FIG. 4 is a functional block diagram showing a conventional image recognition device.

[Explanation of symbols]

10 image recognition device 11 target object 20 image recognition means (multilayer neural network) 22 input layer 23 intermediate layer 24 output layer 25 single density detection unit 25A sigmoid function 30 input means 31 optical lens system 32 video signal amplification processing section 33 Delay line 34 Image input section

Claims (6)

[Claims] 1. An image recognition apparatus for recognizing input image information by using a multilayer neural network, which comprises an input layer, an intermediate layer, and an output layer, as a layer, as image recognition means. An image recognition device, characterized in that a single density detection unit is provided between the intermediate layer and the intermediate layer to correct a variation in illumination when image information is input. 2. The single concentration detecting unit is respectively coupled to each unit constituting the intermediate layer by a weight learned in advance for each unit constituting the intermediate layer, and is learned in advance for each unit constituting the input layer. The image recognition apparatus according to claim 1, wherein the image recognition apparatus is configured to be coupled to each of the units according to the weighted weight. 3. Each unit forming the input layer is connected to each of all the units forming the intermediate layer by a weight learned in advance for all the units forming the intermediate layer, and each unit forming the intermediate layer is formed. The image recognition device according to claim 1 or 2, wherein the unit is connected to each of all the units that constitute the output layer by weights learned in advance. 4. The image recognition apparatus according to claim 1, wherein the input / output function of the single density detection unit is a sigmoid function. 5. The weight between the single concentration detection unit and each unit forming the intermediate layer, or the weight between the single concentration detection unit and each unit forming the input layer is the sigmoid. The image recognition apparatus according to claim 1, wherein the input value to the function is learned in advance so that it falls within a predetermined range. 6. The intermediate layer can correct the illumination variation by changing its own amplification factor in accordance with a change in the output from the single concentration detection unit. The image recognition device according to claim 1.