JPH0836643A - Learning method by neural net in image recognition - Google Patents

Abstract

PURPOSE:To provide a learning method by a neural net in image recognition in which a pattern of an image is flexibly and accurately recognized almost without additional learning by having only to register an image of a standard pattern. CONSTITUTION:An image input device 2 picks up an image of an image pickup object 1 on which a standard character is drawn and the image data are A/D- converted and stored once in a frame memory 3. Based on the image of the standard pattern to be stored, an image (1) with an expanded pattern in which character lines are thickened through expansion, an image (2) of a contracted pattern in which character lines are thinned through contraction, and an image (3) whose characters are reduced in the vertical and horizontal directions are generated and the generated images (1)-(3) are stored in a frame memory 5 for generated image.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method using a neural network in image recognition.

[0002]

2. Description of the Related Art Conventionally, in an image processing apparatus using a learning method using a neural network, all pattern data of an image to be judged as a correct answer is fetched, registered and learned.

[0003]

In the method of taking in all the pattern data of the image to be judged as the correct answer and registering and learning as described above, all the images to be the correct answer must be taken in from the image input device, I couldn't register without that image as a sample. Therefore, complete learning was not completed until the samples were prepared, and every time there was a mistake in judgment, the image was additionally climbed and relearned. However, this method cannot be used for judgment during re-learning, which is a practical problem.

In learning by back-propagation in a neural network, an error from an ideal value is calculated by calculating from an input pattern for learning, and the progress of learning is judged from the numerical value. In other words, we calculated the average of the errors of all input patterns to be learned, and judged that the learning was completed when the average value became small to some extent, but with this method even if only a certain pattern could not be learned, it is averaged. As a result, the average value becomes small and learning may end. In addition, when the completion is judged by the worst value of the error, there is a problem that what is learned well is over-learned.

The present invention has been made in view of the above problems, and an object of the inventions of claims 1 to 3 is to register an image of a standard pattern, with almost no additional learning. It is another object of the present invention to provide a learning method of a neural network in image recognition, which can flexibly and accurately recognize an image pattern. The object of the invention of claim 4 is, in addition to the object of the invention of claim 1, to provide a learning method of a neural network in image recognition, which enables detailed learning to proceed.

The object of the invention of claim 5 is, in addition to the object of the invention of claim 1, to provide a learning method of a neural network in image recognition, which enables learning to proceed at high speed. It is an object of the invention of claim 6 to provide a learning method of a neural network in image recognition, which can surely perform sufficient learning and can automatically determine the completion of learning before over-learning. .

[0007]

In order to achieve the above object, according to the invention of claim 1, before starting the learning work, an image of a predetermined pattern taken by the image input device is registered as a standard of the correct answer pattern, and It is characterized in that an image of a deformation pattern such as an expansion pattern, a contraction pattern, or a compression pattern which is assumed to be an allowable range in image recognition is generated and registered based on the image of the standard pattern, and the generated and registered image is subjected to learning.

According to a second aspect of the invention, in the first aspect of the invention, the predetermined pattern is a character pattern. According to a third aspect of the invention, in the first aspect of the invention, the predetermined pattern is a figure pattern other than a character pattern. The invention of claim 4 is characterized in that, in the invention of claim 1, learning by backpropagation is applied using a sequential correction method for correcting the coupling coefficient for each pattern.

The invention of claim 5 is characterized in that learning by backpropagation is applied by using a collective correction method for collectively correcting the coupling coefficient after calculating all outputs. According to the invention of claim 6, in the learning method by the neural network in the image recognition in which the image registered as the correct answer pattern is learned by the back-propagation, the learning completion time is set based on the learning error and the achievement degree of learning of each output. It is characterized by making a judgment.

[0010]

According to the inventions of claims 1 to 3, before starting the learning work, an image of a predetermined pattern photographed by the image input device is registered as a standard of the correct pattern, and based on the image of the standard pattern. An image of a deformed pattern such as an expansion pattern, a contraction pattern, or a compression pattern, which is assumed to be an allowable range in image recognition, is generated and registered. Therefore, only by registering the image of the standard pattern, the allowable range of the image to be recognized is provided. Therefore, sufficient learning is possible, so that it is possible to flexibly respond to changes in imaging conditions due to lighting or changes in the image to be inspected without additional learning, and to accurately recognize patterns of characters and figures. Moreover, since the need for additional learning is small, the interruption of the pattern recognition operation due to learning can be almost eliminated.

According to the invention of claim 4, in the invention of claim 1, learning is performed by backpropagation by using a sequential correction method for correcting the coupling coefficient for each pattern, so that the learning is advanced in a fine manner. You can According to the invention of claim 5, learning is performed by backpropagation by using a batch correction method for collectively correcting the coupling coefficient after calculation of all outputs, so that the learning can be advanced at high speed.

According to the invention of claim 6, in the learning method by the neural network in the image recognition in which the image registered as the correct answer pattern is learned by the back-propagation, based on the learning error and the achievement level of learning of each output. Since the learning completion time is determined by performing the learning, it is possible to surely perform sufficient learning and automatically complete the learning before over-learning.

[0013]

Embodiments of the present invention will be described below with reference to the drawings. (Embodiment 1) FIG. 2 shows an outline of an image processing apparatus using this embodiment. As shown in FIG. 1, an image pickup target 1 on which standard characters are drawn is picked up by an image input apparatus 2, The image data is temporarily stored in the frame memory 3 after A / D conversion, and an image of an expanded pattern in which the character line is expanded and expanded by the processing operation of the MPU unit 4 based on the stored standard image data, An image of a contraction pattern in which the character line is contracted and the character line is thinned, and an image in which the character itself is contracted vertically and horizontally are generated and registered in the frame memory 5 for the generated image. In addition, 6 in FIG. 2 shows a system bus.

In this case, by providing each of the expansion / contraction ratios in a plurality of stages, it is possible to create a plurality of stages of an image of a standard pattern and an image having a different character thickness from the image obtained from the object 1 as shown in FIG. These are possible assuming that the brightness changes or the binarized image changes due to the difference in quality of the target image itself. Further, in the case of cutting out a character pattern, when noise NS exists in the background as shown in FIG. 4A, when the character pattern cut out affected by the background noise is affected by the noise NS. The part including the part is cut out as a character pattern, and the original character pattern part is shortened as shown in FIG. To deal with this, an image of a pattern that is compressed vertically and horizontally is created. By having a plurality of compression rates, it is possible to create images in various compressed states centering on a standard pattern image. In FIG. 3, 1 dot is packed in the lower part, 1 mark is packed in the upper part, 1 mark is packed in the left part, and 1 mark is packed in the right part.

In this way, by combining the generation methods based on the two factors on the basis of the captured image, it is possible to create a plurality of character patterns having various deformation patterns. Note that FIG. 4C shows the case where there is no noise in the background, and FIG. 4D shows the cutout of the character pattern in that case.
The image data of each pattern thus obtained is stored in the frame memory 5 for the generated image, and is used in the next step of learning work by the neural network.

As shown in FIG. 1, the nutal net has a coupling coefficient (weight) between an input and an output as a basic structure, is composed of an input layer I, an intermediate layer II, and an output layer III, and is given to the input layer I. The image data is propagated to the intermediate layer II by being strengthened and weakened by the coupling coefficient of the first layer, and is propagated to the output layer III by being strengthened and weakened by the coupling coefficient of the second layer from the intermediate layer II. At this time, if an appropriate coupling coefficient is given, a predetermined output value can be obtained and a character pattern can be discriminated (recognized). Further, the work of obtaining an appropriate coupling coefficient for obtaining such a result is called learning, and in this embodiment, learning is realized by backpropagation. This is a method of obtaining the difference between the ideal output state and the actual output value at that time, and adjusting the coupling coefficient so that the error value becomes small. To further explain, when the image of “A” is given as shown in FIG. 5, the output value of No. 1 is large, and other outputs are small, and the image of “B” is given. If the output of No. 2 is large and the other outputs are small, the character can be recognized as a result. The image data is converted into binary data (or grayscale data) and then taken into the input layer. 5A shows the output value, and FIG. 5B shows the ideal value for the "A" image.
(C) shows the respective ideal values for the "B" image.

The methods described above include a sequential correction method for reflecting an error of one pattern in the coupling coefficient each time the error is obtained, and a batch correction method for calculating the errors of all registered patterns and correcting the coupling coefficient at once. There is. The former iterative correction method has a tendency to perform learning finely, but the learning time tends to be long, and the iterative correction method has a tendency to proceed rapidly but at a high speed. In the present embodiment, the MPU unit 4 is programmed so as to select one of these methods according to conditions such as the number of character patterns to be generated and the number of character patterns to be registered.

By reflecting the generated image data obtained by the above method on the input image data used for learning by backpropagation, all the generated images can be discriminated. After the image data is registered, Since the image data used for learning is stored in the frame memory 5, the learning work can be performed only by the image processing device.

Although the above description with reference to FIGS. 1 to 4 is about the registration of a character pattern such as an alphabet and the generation and registration of a modified pattern thereof, as shown in FIGS. 6 to 7, it is also possible to use a graphic pattern other than the character pattern. Registration of a standard pattern and generation and registration of a modified pattern thereof may be performed before the learning work by the same method. Of course, the learning method may be the same method. The hardware configuration of the image processing apparatus used is the same as that shown in FIG.
Further, elements having the same meaning, configuration, or function as those in FIGS. 1 to 4 are given the same numbers and symbols.

Since the character patterns in FIGS. 1 to 4 are replaced by graphic patterns in FIGS. 6 to 8,
A particular description will be given here with reference to the description corresponding to FIGS. (Embodiment 2) By the way, in learning by back-propagation in a neural network, an error from an ideal value is calculated by calculating from an input pattern for learning, and the progress of learning is judged based on the numerical value. In other words, the average of the errors of all input patterns for learning is calculated, and it is judged that learning is completed when the average value becomes small to some extent, but with this method even if only a certain pattern is not learned, it is averaged. As a result, the average value becomes small and learning may end. Further, if the completion is judged by the worst value of the error, what has already been learned well may be overlearned by further learning. It is the invention of claim 6 that solves this problem, and this embodiment corresponds to the invention of claim 6.

In this embodiment, the image processing apparatus shown in FIG. 2 is used as hardware, and the learning process is performed by the MPU unit 4. Then, learning is performed by backpropagation in the above-described neural network shown in FIG. 1 (FIG. 6) or FIG. Thus, in the case of performing the learning work by the back propagation, in this embodiment, the reference value (pass value) of each output for the image of each input pattern is set, and the output number exceeding the reference value is the ratio to the total output number. I am trying to express it. In other words, 0% if any output does not exceed the pass value, and 100% if all exceed the pass value. This ratio is called the learning rate.
Figure 9 shows the relationship between this learning rate (a) and time,
Judging the progress of learning with this proficiency rate will surely allow sufficient learning, and by adding processing that does not learn to those exceeding the passing value, it is possible to prevent the problem of overlearning. When the error becomes large due to the influence of the learning of the pattern and falls below the pass value, if it is made to be the subject of learning again, excessive learning does not occur, and reliable and sufficient learning can be performed,
In addition, it is possible to automatically judge the completion, and it is possible to automate the result learning work.

The curve (B) shown in FIG. 9 indicates a squared error, and this squared error is conventionally learned depending on whether the learner can make a normal judgment empirically or with a test image different from the learning. It was used when determining the completion of.

[0023]

According to the inventions of claims 1 to 3, before starting the learning work, an image of a predetermined pattern imaged by the image input device is registered as a standard of a correct pattern, and based on the image of the standard pattern. An image of a deformed pattern such as an expansion pattern, a contraction pattern, or a compression pattern, which is assumed to be an allowable range in image recognition, is generated and registered. Therefore, only by registering the image of the standard pattern, the allowable range of the image to be recognized is provided. Therefore, sufficient learning is possible, so that it is possible to flexibly respond to changes in imaging conditions such as lighting and changes in the inspection target image without performing additional learning.
In addition, it is possible to accurately recognize a pattern of characters and figures, and since there is little need for additional learning, there is an effect that interruption of the pattern recognition operation due to learning can be almost eliminated.

According to the invention of claim 4, in the invention of claim 1, learning is performed by backpropagation by using a sequential correction method in which the coupling coefficient is corrected for each pattern, so that the learning can be advanced finely. There is an effect. According to the invention of claim 5, learning is performed by backpropagation by using the collective correction method of collectively modifying the coupling coefficient after calculation of all outputs, so that the learning can be advanced at high speed. .

According to a sixth aspect of the present invention, in a learning method by a neural network in image recognition in which an image registered as a correct answer pattern is learned by backpropagation, learning is performed based on a learning error and a degree of achievement of learning of each output. Since the completion time is determined, it is possible to automatically and surely perform sufficient learning and to complete the learning before over-learning, and therefore, it is possible to automate the learning work.

[Brief description of drawings]

FIG. 1 is a flowchart of a process corresponding to a character pattern according to a first embodiment of the present invention.

FIG. 2 is a hardware configuration diagram of an image processing apparatus used in the above.

FIG. 3 is an explanatory diagram of generation of a modified pattern in the above character pattern.

FIG. 4 is an explanatory diagram of cutting out a character pattern of the above.

FIG. 5 is an explanatory diagram of back propagation of the neural network of the above.

FIG. 6 is a flowchart of a process corresponding to the graphic pattern of the above.

FIG. 7 is an explanatory diagram of generation of a modified pattern in the graphic pattern of the above.

FIG. 8 is an explanatory diagram of cutting out the graphic pattern of the above.

FIG. 9 is an explanatory diagram of a relationship between a learning rate and time used in Example 2 of the present invention.

[Explanation of symbols]

 1 Image Target 2 Image Input Device 3, 5 Frame Memory ~ Image

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] December 26, 1994

[Procedure Amendment 1]

[Document name to be amended] Statement

[Name of item to be corrected] Claim 4

[Correction method] Change

[Correction content]

[Procedure Amendment 2]

[Document name to be amended] Statement

[Name of item to be corrected] Claim 5

[Correction method] Change

[Correction content]

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0004

[Correction method] Change

[Correction content]

[0004] In accordance with the learning back pro Page <br/> Deployment in neural network calculates an error between the ideal value calculated from the input pattern applied to the learning, and determining the progress of learning from that number. In other words, we calculated the average of the errors of all input patterns to be learned, and judged that the learning was completed when the average value became small to some extent, but with this method even if only a certain pattern could not be learned, it is averaged. As a result, the average value becomes small and learning may end. In addition, when the completion is judged by the worst value of the error, there is a problem that what is learned well is over-learned.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0007

[Correction method] Change

[Correction content]

[0007]

Means for Solving the Problems] In the invention of claim 1 in order to achieve the above object, before entering the learning task to register the image of the predetermined path terpolymers emissions captured by the image input device as a standard correct pattern Based on the image of the standard pattern, an image of a deformation pattern such as an expansion pattern, a contraction pattern, or a compression pattern which is assumed to be an allowable range in image recognition is generated and registered, and the generated and registered image is subjected to learning. To do.

[Procedure Amendment 5]

[Document name to be amended] Statement

[Correction target item name] 0008

[Correction method] Change

[Correction content]

According to a second aspect of the invention, in the first aspect of the invention, the predetermined pattern is a character pattern. According to a third aspect of the invention, in the first aspect of the invention, the predetermined pattern is a figure pattern other than a character pattern. In the invention of claim 4, in the invention of claim 1, wherein the applying learning by back pro Page Shon using successive correction method Modifications to the coupling coefficient for each pattern.

[Procedure correction 6]

[Document name to be amended] Statement

[Correction target item name] 0011

[Correction method] Change

[Correction content]

According to the invention of claim 4, in the invention of claim 1, since multiplying learning by back pro Page Shon using successive correction method Modifications to the coupling coefficient for each pattern, thereby advancing the finely learning be able to.
According to the invention of claim 5, since applying learning by the backplane <br/> b Page Shon using bulk modify methods Modifications to the coupling coefficient collectively after calculation of all output, to advance learning speed be able to.

[Procedure Amendment 7]

[Document name to be amended] Statement

[Correction target item name] 0013

[Correction method] Change

[Correction content]

[0013]

Embodiments of the present invention will be described below with reference to the drawings. (Embodiment 1) FIG. 2 shows an outline of an image processing apparatus using this embodiment. As shown in FIG. 1, an image pickup target 1 on which standard characters are drawn is picked up by an image input apparatus 2, and stores the image data in the a / D conversion after temporarily frame memory 3, the processing operation of the MPU unit 4 the data in this storage the standard image based, expanded in the expansion path terpolymers down character line becomes thicker An image, an image of a contraction pattern in which a character line is contracted and the character line is thinned, and an image in which the character itself is contracted vertically and horizontally are generated and registered in the frame memory 5 for the generated image. In addition, 6 in FIG. 2 shows a system bus.

[Procedure Amendment 8]

[Document name to be amended] Statement

[Correction target item name] 0020

[Correction method] Change

[Correction content]

Since the character patterns in FIGS. 1 to 4 are replaced by graphic patterns in FIGS. 6 to 8,
A particular description will be given here with reference to the description corresponding to FIGS. (Embodiment 2) By the way, back processing by a neural network
In accordance learning Page Deployment calculates an error between the ideal value calculated from the input pattern applied to the learning, and determining the progress of learning from that number. In other words, the average of the errors of all input patterns for learning is calculated, and it is judged that learning is completed when the average value becomes small to some extent, but with this method even if only a certain pattern is not learned, it is averaged. As a result, the average value becomes small and learning may end. Further, if the completion is judged by the worst value of the error, what has already been learned well may be overlearned by further learning. It is the invention of claim 6 that solves this problem.
Corresponding to the invention of.

[Procedure Amendment 9]

[Document name to be amended] Statement

[Correction target item name] 0021

[Correction method] Change

[Correction content]

In this embodiment, the image processing apparatus shown in FIG. 2 is used as hardware, and the learning process is performed by the MPU unit 4. The learning back professional path in the above-described neural network shown in FIG. 1 (FIG. 6) or 5
Performs learning by the gate Shon. Thus, in the case of performing the learning work by the back propagation, in this embodiment, the reference value (pass value) of each output for the image of each input pattern is set, and the output number exceeding the reference value is the ratio to the total output number. I am trying to express it. In other words, 0% if any output does not exceed the pass value, and 100% if all exceed the pass value. This ratio is called the learning rate. FIG. 9 shows the relationship between this learning rate (a) and time. If the learning progress is judged by this learning rate, sufficient learning can be surely performed, and if the learning rate exceeds the passing value. If you add processing that does not learn, you can prevent the problem of excessive learning, and if the error becomes large due to the influence of the learning of other patterns and it falls below the pass value, it will be over-learned if you make it the object of learning again. And surely enough learning can be done,
In addition, it is possible to automatically judge the completion, and it is possible to automate the result learning work.

[Procedure Amendment 10]

[Document name to be amended] Statement

[Name of item to be corrected] 0024

[Correction method] Change

[Correction content]

The invention of claim 4 is the invention of claim 1, since multiplying learning by back pro Page Shon using successive correction method Modifications to the coupling coefficient for each pattern, be allowed to proceed to finely learning The effect is that you can do it. The invention of claim 5, the effect of so applying learning by back pro Page Shon using bulk modify methods Modifications to the coupling coefficient collectively after calculation of all output, it can proceed learn faster is there.

Claims (6)

[Claims] 1. Before starting a learning operation, an image of a predetermined pattern captured by an image input device is registered as a standard of a correct pattern, and an expansion pattern assumed to be an allowable range in image recognition based on the image of the standard pattern. A learning method using a neural network in image recognition, wherein images of deformation patterns such as contraction patterns and compression patterns are generated and registered, and the generated and registered images are subjected to learning. 2. The learning method using a neural network in image recognition according to claim 1, wherein the predetermined pattern is a character pattern. 3. The learning method by a neural network in image recognition according to claim 1, wherein the predetermined pattern is a graphic pattern other than a character pattern. 4. A learning method by a neural network in image recognition according to claim 1, wherein learning by backpropagation is performed by using a sequential correction method for correcting the coupling coefficient for each pattern. 5. The learning by a neural network in image recognition according to claim 1, wherein learning by backpropagation is performed by using a batch correction method for collectively correcting the coupling coefficient after calculating all outputs. Method. 6. In a learning method using a neural network in image recognition in which an image registered as a correct answer pattern is learned by back-propagation, the learning completion time is determined based on the learning error and the achievement level of learning of each output. A learning method using a neural network in image recognition.