Engraved Character Reading System and Engraved Character Reading Method

Abstract

To provide a system for reading an engraved character and a method for reading an engraved character, capable of reading an engraved character with high accuracy.SOLUTION: A system 1 for reading an engraved character includes character position identification means 20 that identifies a position of an engraved character for each character in a photographic image obtained by imaging means 10, and character type determination means 30 that determines which of candidate characters a character type is in an engraved character image for each character at the position identified by the character position identification means 20. The character position identification means 20 identifies the position on the basis of a position learning model 21 obtained by learning a feature of an area where the engraved character is located for each character of each candidate character type by machine learning. The character type determination means 30 determines the character type on the basis of a position learning model 31 obtained by learning a feature of a shape for each character of each candidate character type by machine learning.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an engraved character reading system and an engraved character reading method for reading engraved characters.

Background Art

[0002] In an engine assembly line of a vehicle, for example, numbers of metal assembly instructions engraved on an engine block are read, and metal numbers are selected based on the numbers. The reading of the numbers is performed, for example, visually, but in recent years, automatic character reading technology has been developed. For example, Patent Document 1 describes a character recognition device that reads a digital image from a printed surface on which an impressed number sequence is printed by a CCD camera, transfers it to a computer, moves a slit narrower than the width of each impressed number, extracts a slit image from the digital image, and recognizes the character type using a neural network. Further, Patent Document 2 describes an accounting processing system that performs character recognition from a read image by OCR processing and generates text data.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, with these image analysis techniques, in the case of printed or handwritten characters where there is a color difference between the characters and the background, the characters can be recognized. However, in the case of engraved characters where the characters and the background are the same color, problems such as gloss and dirt on the surface have an adverse effect during image analysis, and there is a problem that they may not be recognized as characters. In particular, if there are double impressions or misalignments of the characters, for example, as shown in FIG. 4, even when trying to obtain the outline of the characters by binarization processing, gloss, dirt, or minute irregularities on the skin are picked up as noise, resulting in a low discrimination rate and difficulty in recognizing the characters.

[0005] The present invention has been made based on such problems, and an object thereof is to provide an engraved character reading system and an engraved character reading method capable of reading engraved characters with high accuracy.

Means for Solving the Problems

[0006] The engraved character reading system of the present invention reads which of a plurality of candidate character types the engraved character to be read belongs to, and includes a photographing means for photographing an engraved surface provided with the engraved character, a character position specifying means for specifying the position of the engraved character for each character from the photographed image obtained by the photographing means, and a character type discriminating means for discriminating which of the candidate character types the character type of the engraved character image for each character at the position specified by the character position specifying means in the photographed image obtained by the photographing means is. The character position specifying means specifies based on a position learning model in which the characteristics of the region where the engraved character is located for each character are learned for each candidate character type by machine learning, and the character type discriminating means discriminates based on a character type learning model in which the characteristics of the shape of the engraved character for each character are learned for each candidate character type by machine learning.

[0007] The method for reading engraved characters of the present invention reads which of a plurality of candidate character types the engraved characters to be read belong to, and includes a photographing procedure for photographing an engraved surface on which the engraved characters are provided, and for the photographed image obtained by the photographing means, a character position specifying procedure for specifying the position of the engraved characters one by one, and for each engraved character image at the position specified by the character position specifying procedure among the photographed images obtained by the photographing procedure, a character type discrimination procedure for discriminating which of the candidate character types the character type is. In the character position specifying procedure, the position is specified using a position learning model in which the characteristics of the region where the engraved characters are located for each character are learned for each candidate character type by machine learning. In the character type discrimination procedure, the discrimination is performed using a character type learning model in which the characteristics of the shape of the engraved characters for each character are learned for each candidate character type by machine learning.

Advantages of the Invention

[0008] According to the present invention, using a position learning model in which the characteristics of the region where the engraved characters are located for each character are learned for each candidate character type by machine learning, the position of the engraved characters is specified for each character, and using a character type learning model in which the characteristics of the shape of the engraved characters for each character are learned for each candidate character type by machine learning, it is determined which of the candidate character types the character type of the engraved character image at the specified position is. Therefore, it is possible to detect and learn the characteristics including the gloss and dirt on the surface, and the engraved characters can be read with high accuracy.

Brief Description of the Drawings

[0009]

Figure 1

Figure 2

Figure 3

Figure 4

Embodiments for Carrying Out the Invention

[0010] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0011] FIG. 1 shows the configuration of an engraved character reading system 1 according to an embodiment of the present invention. This engraved character reading system 1 reads which of a plurality of candidate character types the engraved character to be read belongs to. The engraved character is an engraved character formed by stamping. The object to be read may be one character or a plurality of characters. In the present embodiment, a case where engraved characters provided on a workpiece M such as an engine block are read on a production line such as an engine assembly line will be described as an example.

[0012] The engraved character reading system 1 includes, for example, a photographing means 10 that photographs an engraved surface on which an engraved character to be read is provided, a character position specifying means 20 that specifies the position of the engraved character for each character in the photographed image obtained by the photographing means 10, and, for each engraved character image for each character at the position specified by the character position specifying means 20 in the photographed image obtained by the photographing means 10, a character type discriminating means 30 that discriminates which of the candidate character types the character type is.

[0013] The photographing means 10 is constituted by, for example, a camera such as a CCD camera, and is fixedly arranged so as to photograph the engraved characters of the workpiece M conveyed on the production line. The photographing range of the photographing means 10 is preferably set so as to include all of the objects to be read, for example.

[0014] The character position specifying means 20 can be configured by, for example, a computer and is configured to function as the character position specifying means 20 by executing a program. The character position specifying means 20 is preferably connected to the photographing means 10, for example, and is configured to specify the position of each printed character from the photographed image obtained by the photographing means 10 and cut out and save the printed character image for each character at the specified position. The specification of the position of the printed character is performed, for example, based on a position learning model 21 in which the characteristics of the region where the printed character is located for each character are learned for each candidate character type by machine learning.

[0015] Specifically, the character position specifying means 20 includes, for example, a position learning model 21 in which the characteristics of the region where the printed character is located for each character are learned for each candidate character type by machine learning, a detection unit 22 that detects each printed character from the photographed image using the position learning model 21 and cuts it out as a printed character image, and a storage unit 23 that stores the printed character image cut out by the detection unit 22 together with position information. Further, the character position specifying means 20 preferably has, for example, a position learning model generation means 24 for generating the position learning model 21.

[0016] The position learning model 21 can be generated by the position learning model generation means 24. The position learning model generation means 24, for example, takes a captured image of a printed surface provided with printed characters for a plurality of types of candidate character types, and for each character, surrounds the printed character with a frame or the like and sets the image within the designated area as the position learning image. It is preferably configured to extract the feature amount of the position learning image by deep learning (Deep learning) and learn it as the feature amount of the position where any one of the candidate character types of printed characters is located. For example, when the candidate character types are 7 types of numbers from 0 to 6, from the captured image of the printed surface provided with 7 types of printed characters from 0 to 6, the image within the area surrounded by a frame or the like for each character is set as the position learning image, and it is preferably learned as the position where any one of the printed characters is located. The printed characters of each candidate character type may be provided on the same printed surface and captured, or may be provided on different printed surfaces and captured. For deep learning, it is preferable to use a convolutional neural network (CNN).

[0017] As a result, the position learning model 21 has detected and learned the combined feature amounts of each candidate character type for the area where one printed character is located, including the gloss and dirt on the surface of the printed surface. That is, the character position specifying means 20 is configured to be able to specify the positions of the printed characters of a plurality of types of candidate character types by one position learning model 21. Note that as the position learning image, it is preferable to include various modes that may be printed characters to be read, particularly in view of the range of variations in parts flowing on the corresponding production line, such as images of double-struck characters, images of characters with rotated orientations such as obliquely or upside down, and images of characters with different brightnesses.

[0018] The character discrimination means 30 can be configured by, for example, a computer and is configured to function as the character discrimination means 30 by executing a program. The character discrimination means 30 is connected to, for example, the character position specifying means 20, and based on the printed character image at the position specified by the character position specifying means 20 and the character type learning model 31 in which the characteristics of the shape of the printed characters are learned for each candidate character type by machine learning, it is configured to discriminate the character type.

[0019] Specifically, the character discrimination means 30 includes, for example, a character type learning model 31 in which the characteristics of the shape of the printed characters are learned for each candidate character type by machine learning, a determination unit 32 that inputs the printed character image into the character type learning model 31 and determines which of the candidate character types the character type is based on the output value obtained thereby, and a display unit 33 such as a display that displays the character type determined by the determination unit 32. Further, the character determination means 30 preferably has, for example, a character type learning model generation means 34 that generates the position type learning model 31.

[0020] The character type learning model 31 can be generated by the character type learning model generation means 34. The character type learning model generation means 34, for example, uses an image within a region specified by surrounding each printed character with a frame or the like one by one from a photographed image of a printed surface provided with printed characters for a plurality of types of candidate character types as a character type learning image, extracts the feature amount of the character type learning image for each candidate character type by deep learning, and preferably associates the feature amount with the candidate character type of the correct character type for learning. For example, when the candidate character types are 7 types of numbers from 0 to 6, it is preferable to use an image within a region specified by surrounding each printed character with a frame or the like one by one from a photographed image of a printed surface provided with 7 types of printed characters from 0 to 6 as a character type learning image and associate it with the candidate character type of the correct character type for learning. It is preferable to use the same image as the position learning image for the character type learning image. It is preferable to use a convolutional neural network for deep learning.

[0021] As a result, the character type learning model 31 has detected and learned the feature amounts of the shapes of the printed characters for each of a plurality of candidate character types, including the gloss, dirt, etc. on the surface of the printing surface. That is, the character discrimination means 30 is configured such that one character type learning model 31 can discriminate which of the candidate character types the character type is. Note that, similar to the position learning image, the character type learning image preferably includes various modes that may be the printed characters to be read, particularly in view of the range of variations in the parts flowing on the corresponding production line. For example, it includes images of double-struck characters, images of characters whose orientation is rotated such as obliquely or upside down, and images of characters with different brightness levels.

[0022] The determination unit 32 is connected to, for example, the storage unit 24, the character type learning model 31, and the display unit 33, respectively. The discrimination unit 32 is preferably configured to input the printed character image stored in the storage unit 24 into the character type learning model 31 and determine, based on the output value obtained from the character type learning model 31, for example, the score value of each candidate character type, that the candidate character type with the highest degree of coincidence is the one. Note that, when there is no candidate character type with a high degree of coincidence or when there are two or more candidate character types with the same degree of coincidence, it may be determined that the character type is unknown.

[0023] The display unit 33 is preferably configured to display the printed characters on the printing surface based on, for example, the character type of the printed character image determined by the determination unit 32 and the position information of the printed character image.

[0024] FIG. 2 shows an example of the hardware configuration of the character position specifying means 20 and the character type discriminating means 30. The character position specifying means 20 and the character type discriminating means 30 include, for example, a CPU (Central Processing Unit) 41, a ROM (Read Only Memory) 42, a RAM (Random Access Memory) 43, an HDD (hard disk drive) 44, and an operation interface (operation I / F) 45. The CPU 41 executes various processes according to various programs recorded in the ROM 42 or various programs loaded from the HDD 44 into the RAM 43. The RAM 43 appropriately stores data and the like necessary for the CPU 41 to execute various processes. The HDD 44 stores various data.

[0025] This engraved character reading system 1 is used, for example, as follows. FIG. 3 shows the flow of an engraved character reading method using the engraved character reading system 1. First, as a preparation procedure, a position learning model 21 is generated by the position learning model generation means 24, and a character type learning model 31 is generated by the character type learning model generation means 34 (preparation procedure; step S110).

[0026] Specifically, for example, an engraving surface provided with engraved characters for a plurality of types of candidate character types is photographed by the photographing means 10, and from the obtained photographed image, the image within the region specified by surrounding each engraved character with a frame or the like is used as a position learning image, and the feature amount of the position learning image is extracted by deep learning, and a position learning model 21 is generated by learning it as the feature amount of the position where any one of the supplementary character types is located. Further, for example, the same image as the position learning image is used as a character type learning image, and the feature amount of the character type learning image is extracted for each candidate character type by deep learning, and a character type learning model 31 is generated by associating the feature amount with the candidate character type of the correct character type. For example, when the candidate character types are 7 types of numbers from 0 to 6, from the photographed image of the engraving surface provided with 7 types of engraved characters from 0 to 6, the image within the region specified by surrounding each character with a frame or the like is prepared as a character position learning image and a character type learning image.

[0027] Next, for example, in a production line such as an engine assembly line, the engraved characters provided on the workpiece M such as the engine block are read as follows. First, the imaging means 10 images the engraved surface of the workpiece M where the engraved characters, for example, the numbers "32232" are provided (imaging procedure; step S121). Next, for example, the character position specifying means 20 uses the position learning model 21 to characterize the position of each engraved character in the captured image obtained by the imaging means 10 (character position specifying procedure; step S122). Specifically, for example, the detection unit 22 uses the position learning model 21 to detect each engraved character from the captured image one by one, cuts it out as an engraved character image, and saves the cut-out engraved character image together with the position information in the storage unit 23.

[0028] Subsequently, for example, the character discrimination means 30 uses the character type learning model 31 to discriminate which character type among the candidate character types each engraved character image at the position specified by the character position specifying means 20 belongs to (character type discrimination procedure; step S123). Specifically, for example, the determination unit 32 inputs the engraved character image into the character type learning model 31, and based on the output value obtained from the character type learning model 31, determines that it is the character type with the highest degree of coincidence among the candidate character types, and displays it on the display unit 33. For example, when the score values of each candidate character type obtained from the character type learning model 31 are 0% for "0", 0% for "1", 98% for "2", 1% for "3", 0% for "4", 1% for "5", and 0% for "6", it is determined that the number is 2 with the highest score value. This is done for each engraved character image, and on the display unit 33, the determined character type is displayed, for example, as "32232" based on the position information of each engraved character image.

[0029] In addition, in the engine assembly line, when the engraved characters provided on the engine block were read by the engraved character reading system 1 and OCR, the reading rate by OCR was 98%, while the reading rate by the engraved character reading system 1 was 100%.

[0030] As described above, according to this embodiment, using the position learning model 21 in which the characteristics of the area where the printed characters are located for each character are learned by machine learning for each candidate character type, the position of the printed characters is specified for each character, and using the character type learning model 31 in which the characteristics of the shape of the printed characters are learned for each candidate character type by machine learning for each character, it is possible to determine which of the candidate character types the character type of the printed character image at the specified position is. Therefore, it is possible to detect and learn the characteristics including the gloss and dirt on the surface, and the printed characters can be read with high accuracy.

[0031] As described above, the present invention has been described with reference to the embodiments. However, the present invention is not limited to the above embodiments and can be variously modified. For example, in the above embodiments, each component has been specifically described. However, the specific structure and shape of each component may be different, and not all of the above-described components need to be provided, or other components may be provided.

[0032] Also, in the above embodiment, the case of reading the printed characters provided on the engine block in the engine assembly line has been specifically described. However, it can also be applied to the case of reading the printed characters in other production lines. Furthermore, in the above embodiment, the case where the candidate character type is a number has been specifically described. However, it can also be applied to character types other than numbers.

Explanation of Reference Numerals

[0033] 1... Printed character reading system, 10... Photographing means, 20... Character position specifying means, 21... Position learning model, 22... Detection unit, 23... Storage unit, 24... Position learning model generation means, 30... Character type discrimination means, 31... Character type learning model 31, 32... Determination unit, 33... Display unit, 34... Character type learning model generation means, 41... CPU, 42... ROM, 43... RAM, 44... HDD, 45... Operation interface

Claims

1. A stamped character reading system for reading which of a plurality of candidate character types the stamped character to be read belongs to, comprising: imaging means for imaging a stamped surface on which the stamped character is provided; character position specifying means for specifying the position of the stamped character for each character from the captured image obtained by the imaging means; character type discriminating means for discriminating which of the candidate character types the character type of each stamped character image at the position specified by the character position specifying means in the captured image obtained by the imaging means is; the character position specifying means specifies based on a position learning model obtained by learning, for each candidate character type, the features of the region where the stamped character is located for each character by machine learning; the character type discriminating means discriminates based on a character type learning model obtained by learning, for each candidate character type, the features of the shape of the stamped character for each character by machine learning; the character type learning model is generated by taking, for each candidate character type, an image within a region designated by surrounding the stamped character for each character from a captured image of the stamped surface on which the stamped character is provided as a character type learning image, extracting the feature amounts of the character type learning images for each candidate character type, and associating the feature amounts with the candidate character types of the correct character types and performing learning. A stamped character reading system characterized by the above.

2. A stamped character reading method for reading which of a plurality of candidate character types the stamped character to be read belongs to, comprising: an imaging procedure for imaging a stamped surface on which the stamped character is provided; a character position specifying procedure for specifying the position of the stamped character for each character from the captured image obtained by the imaging means; a character type discriminating procedure for discriminating which of the candidate character types the character type of each stamped character image at the position specified by the character position specifying procedure in the captured image obtained by the imaging procedure is; in the character position specifying procedure, it is specified using a position learning model obtained by learning, for each candidate character type, the features of the region where the stamped character is located for each character by machine learning; in the character type discriminating procedure, it is discriminated using a character type learning model obtained by learning, for each candidate character type, the features of the shape of the stamped character for each character by machine learning. The character type learning model captures an image of the engraved surface provided with the engraved characters for each candidate character type, and uses the image within the area specified by surrounding the engraved characters one by one from the captured image as a character type learning image. For each candidate character type, it extracts the feature amount of the character type learning image, associates the feature amount with the candidate character type of the correct character type, and generates it by learning. A method for reading engraved characters, characterized by the above.

Images