JP6430672B1 - Reading system and reading method - Google Patents

Abstract

Provided are a reading system and a reading method capable of automatically recognizing only a predetermined identification character string with high accuracy even when imaging conditions are bad or a plurality of character strings exist in an image.
A reading system 40 for reading an identification character string 20 composed of a characteristic character string 21 having a known characteristic and an inspection character 22 from an input image 14 includes the character string as a recognition unit. A character string recognition unit 41 for recognizing the characteristic character string 21 based on a known characteristic from among a plurality of character strings included in the character string, and a characteristic character string 21 searched by the character string recognition unit 41 using a single character as a recognition unit. When the single character recognition unit 42 that recognizes the check character 22 based on the position and the calculated character 25 calculated from the characteristic character string 21 recognized by the character string recognition unit 41 match the check character 22, the identification character string And an output unit 44 that recognizes and outputs as 20.
[Selection] Figure 4

Description

  The present invention relates to a reading system and a reading method, and specifically to a reading system and a reading method that automatically recognizes a predetermined identification character string with high accuracy.

  As an apparatus for automatically recognizing a character string for identification, binarization processing is performed on an image captured by a camera, and recognition processing is performed character by character for a binarized image that matches a feature amount of a symbol. Devices have been proposed that can be recognized and recognized as identification character strings. (For example, refer to Patent Document 1).

JP-A-10-105647

  By the way, as a technique for recognizing characters, there is an attempt to improve identification accuracy using a neural network. That is, in the apparatus described in Patent Document 1, it is possible to further improve the identification accuracy by using a neural network for character recognition.

  As a typical neural network for recognizing characters, a convolutional neural network represented by LeNet is exemplified. The convolutional neural network is a neural network that recognizes a single character by learning to classify an image in which the single character is reflected into a plurality of categories.

  However, in the convolutional neural network, since the recognition unit is limited to a single character, in order to recognize the identification character string, it is necessary to perform image processing to cut out individual single characters as a target from the image. Therefore, depending on the imaging conditions such as the influence of the shadow reflected in the image and the contrast ratio in the image, there is a problem that the recognition accuracy of the identification character string recorded on the surface of the article is lowered and the recognition rate is lowered. . Further, when there are a plurality of character strings in the image, it is difficult to recognize only the identification character string to be read, and it is necessary to perform processing on unnecessary character strings.

  An object of the present invention is to provide a reading system and a reading method capable of automatically recognizing only a predetermined identification character string with high accuracy even when imaging conditions are bad or a plurality of character strings exist in an image. It is.

The reading system of the present invention that achieves the above-mentioned object is calculated by applying a predetermined rule to a feature character string having a known feature and a feature character string arranged in the vicinity of the feature character string. A reading system for reading an identification character string consisting of a single inspection character from an input image, and data in which a character string area occupied by the character string included in the image is detected is input. The feature character string based on the known features from among a plurality of character strings included in the image, with a character string as a recognition unit from a series of feature amounts from the leading end to the tail in the character region A character string recognizing unit for recognizing a single character image data cut out so that a single character included in the image can be discriminated , and the character string as a recognition unit from the single character image data. Recognition part A single character recognition unit that recognizes the inspection character based on the recognized position of the characteristic character string, and a calculated character calculated by applying the predetermined rule to the characteristic character string recognized by the character string recognition unit And an output unit that recognizes and outputs the characteristic character string and the character string in which the check characters are arranged in order as the character string for identification when the check character matches.

The reading method of the present invention that achieves the above object is calculated by applying a predetermined rule to a characteristic character string having a known characteristic and a characteristic character string that is arranged in the vicinity of the characteristic character string. A reading method for reading a character string for identification consisting of a single character inspection character from an input image , between a leading end and a trailing end in a character string region occupied by the character string included in the image A single character included in the image by recognizing the feature character string based on the known feature from among a plurality of character strings included in the image, using a character string as a recognition unit from a series of feature amounts. the but as recognition units of single characters from the single character image data cut out to be discriminated, with respect to the characteristic character string recognizing the check character was recognized based on the position of the feature strings began to probe Applying the prescribed rules When the calculated character and the check character are determined to be calculated, and when it is determined that the calculated character and the check character match, the characteristic character string and the check character are calculated. A character string in which characters are arranged in order is recognized as the identification character string.

  The present invention relates to a character string recognition unit that recognizes a characteristic character string using a known feature as a clue using a character string as a recognition unit, and a single character whose position is specified by recognizing the characteristic character string using a single character as a recognition unit Is used in combination with a single character recognition unit that recognizes as a check character. Therefore, according to the present invention, it is possible to omit the image processing for cutting out the target characters from the image one by one, and to eliminate the decrease in the recognition of the characteristic character string caused by the image processing. At the same time, it is possible to recognize the inspection character with high accuracy by identifying the position of the inspection character by recognizing the characteristic character string, and it is possible to eliminate the uncertain recognition result of the characteristic character string. Thereby, even when the imaging conditions are bad or there are a plurality of character strings in the image, only a predetermined identification character string can be automatically recognized with high accuracy.

It is a block diagram which illustrates embodiment of a reading system. FIG. 2 is an explanatory diagram illustrating an image captured by the imaging apparatus of FIG. 1. It is explanatory drawing which illustrates the arrangement | sequence of the character string for identification. FIG. 2 is a block diagram illustrating the reading system of FIG. 1. It is a part of flowchart which illustrates the reading method. FIG. 6 is a part of a flowchart illustrating a reading method continued from A in FIG. 5. FIG. 7 is a part of a flowchart illustrating a reading method continued from B of FIG. 6. FIG. 8 is a part of a flowchart illustrating a reading method continued from C in FIG. 7. FIG. 8 is a part of a flowchart illustrating a reading method continued from D of FIG. 7. It is explanatory drawing which illustrates the rule of FIG. It is explanatory drawing which illustrates the rule of FIG. It is explanatory drawing which illustrates the rule of FIG.

  Hereinafter, embodiments of the reading system will be described. In this specification, the writing direction of the character string is left horizontal writing in which the character string is arranged in order from left to right, the beginning of the character string is the left end of the character string, and the tail is the right end of the character string.

  As illustrated in FIG. 1, the reading system 40 is a system for recognizing the identification character string 20 described in the cargo 12 handled at the container terminal 11, and the identification character from the image 14 captured by the imaging device 13. This is a system that recognizes the column 20. The reading system 40 is a system incorporated in the edge device 30.

The container terminal 11 is a port facility serving as a node between the sea transportation and the land transportation of the cargo 12. At the container terminal 11, when entering or leaving the container terminal 11 from the gate 15 which is an entrance / exit, an operation of recognizing the identification character string 20 described in the cargo 12 from the image 14 captured by the imaging device 13 is performed. .

  The cargo 12 is exemplified by a container and a container tank. An identification character string 20 is written on the surface of the foreign chassis 16 facing the backward direction in a state where the cargo 12 is mounted on the foreign chassis 16 that goes back and forth between the container terminal 11 and the outside. In addition, the description place of the identification character string 20 is not particularly limited, and may be described on the side surface or the upper surface of the cargo 12.

  The imaging device 13 is a device that captures the image 14 including the identification character string 20 described in the cargo 12 installed in the gate 15 and mounted on the foreign chassis 16. The imaging device 13 is not limited as long as the identification image 20 is included in the captured image 14 and the installation location is not limited. In addition, when the location where the identification character string 20 is described in the cargo 12 can be specified, the imaging device 13 is caused to capture only the specified location in the cargo 12 and is different from the identification character string 20 in the image 14. It is good to reduce the content of the character string.

  As illustrated in FIG. 2, when the cargo 12 is a container, an opening / closing door 17 is formed on the surface thereof, and an identification character string 20 is described in an upper right area of the opening / closing door 17 in a rear view of the foreign chassis 16. The Depending on the arrangement position of the identification character string 20, the identification character string 20 may be divided into a plurality of parts so as to avoid the opening / closing bar 18 of the opening / closing door 17.

  The identification character string 20 is a character string that is likely to be written in the vicinity of the rear corner of the cargo 12 among a plurality of character strings written on the rear surface of the cargo 12, and the cargo 12 is a container. Is likely to be written in the upper right on the back. In addition, the identification character string 20 may be written in the lower right on the back when the cargo 12 is a container tank.

  The identification character string 20 is an eleven-digit character string having a characteristic character string 21 and a check character (also referred to as check digit) 22. The characteristic character string 21 is arranged at the head of the identification character string 20 and includes a first character string 23 and a second character string 24.

  The first character string 23 is arranged at the head of the characteristic character string 21 and is composed of a four-digit alphabet. In the first character string 23, the first three-digit alphabet indicates an owner code that is different for each owner of the cargo 12, and the last one-digit alphabet indicates a device classification identifier. “U” is used as the device classification identifier of the cargo 12. The second character string 24 is arranged at the tail of the characteristic character string 21, is composed of six digits, and indicates a unique serial number of the cargo 12. The inspection character 22 is arranged so as to be adjacent to the identification character string 20, and is composed of single-character numbers calculated by applying a predetermined calculation formula to the characteristic character string 21.

  The owner code is B.B. I. It is a code registered in C (The Bureau International des Containers et du Transport International).

  That is, the identification character string 20 of the cargo 12 is known to include a three-digit alphabet, a one-digit alphabet “U”, a six-digit number, and a one-digit number in order from the head to the tail. It has the characteristics of. In addition, the first character string 23 is B.B among the three-digit alphabet combinations. I. It has a known feature that candidates are narrowed down to owner codes registered in C. Further, the inspection character 22 has a known feature that it is a single digit number arranged near the tail of the feature character string 21. Note that the inspection character 22 may be surrounded by a frame so as to be distinguishable from the characteristic character string 21. Therefore, it may be a known feature that a single-digit number arranged in the vicinity of the feature character string 21 is surrounded by a frame.

  As illustrated in FIGS. 3A to 3C, in the identification character string 20, the arrangement of the first character string 23, the second character string 24, and the inspection character 22 is different for each cargo 12. For example, as illustrated in FIG. 3A, below the first character string 23, the second character string 24 is divided into two character strings of a four-digit number and a two-digit number and arranged in one line. There is a case. Further, as illustrated in FIG. 3B, the second character string 24 is divided into two character strings of a three-digit number and a three-digit number and arranged in one line on the right side of the first character string 23. There is also a case. Further, as illustrated in FIG. 3C, the second character string 24 is divided into two upper and lower columns divided into two character strings of a three-digit number and a three-digit number on the right side of the first character string 23. Sometimes it is arranged. However, in any case, the second character string 24 is arranged in the vicinity of the first character string 23, and the check character 22 is arranged in the vicinity of the tail of the second character string 24.

  As illustrated in FIG. 4, the edge device 30 is installed in the gate 15 and performs various information processing, an internal storage device that can read and write programs and information processing results used to perform the various information processing, and This hardware consists of various interfaces. The edge device 30 is communicably connected to the imaging device 13, the display device 31, and the input device 32, and includes a character string region detection system 33 and a reading system 40 as functional elements. Each functional element is stored in the internal storage device as an independent program, and is executed by the CPU in a timely manner. In addition, as each function element, you may be comprised with the programmable controller (PLC) which each functions independently besides a program.

  The display device 31 and the input device 32 are devices incorporated in a portable terminal that is carried by an operator working at the gate 15. The display device 31 and the input device 32 may be provided in a management building (not shown) in which the edge device 30 is installed in addition to the mobile terminal.

  The character string area detection system 33 receives the image 14 captured by the imaging device 13 and outputs data 34 in which the character string area (also referred to as scene text) Tn and the position Pn of the character string area Tn are detected. It is. The character string area Tn is an area occupied by a character string included in the image 14, and is set as existing in the same area when the interval between adjacent characters is equal to or less than a predetermined threshold. Area. That is, the length in the X direction of the character string region Tn is set according to the number of consecutive characters, and the length in the Y direction is set according to the size of the characters. The position Pn is the position of the character string region Tn in the image 14, and is indicated as a two-dimensional coordinate of a predetermined point (center point of the region in the embodiment) of the character string region Tn with the center of the image 14 as the origin. Note that the character string region Tn can be freely controlled to be enlarged or reduced in the X direction by setting the contrast ratio between the character and the background and the threshold value between the adjacent characters in the character string. It is desirable to be able to change the number of characters in the character string existing in the. Various methods have been proposed for the detection of the character string region Tn. The character string region detection system 33 may be any known method that has been proposed.

  The reading system 40 outputs the identification character string 20 when the data 34 is input and the recognition character string 20 is successfully recognized. On the other hand, the reading system 40 recognizes the recognition character string 20 when the recognition character string 20 fails to be recognized. It is a functional element that outputs the failure. The reading system 40 includes a character string recognition unit 41, a single character recognition unit 42, a determination unit 43, an output unit 44, and an image processing unit 47 as functional elements.

The character string recognition unit 41 is a functional element that receives the data 34 and outputs the recognized characteristic character string 21 to the determination unit 43 and also outputs the position of the characteristic character string 21 to the image processing unit 47. The character string recognizing unit 41 recognizes the characteristic character string 21 by using a complex neural network based on the above-mentioned known features from among a plurality of character strings existing in the data 34 with the character string as a recognition unit. It is a functional element. The character string recognition unit 41 includes a first character string recognition unit 35 and a second character string recognition unit 36.

  The character string recognition unit 41 is a functional element to which single character image data 48 generated by dividing a character string region Tn in which a recognizable characteristic character string 21 exists for each character is input as teacher data.

  As the composite neural network, a neural network combined with a convolutional neural network (CNN) and a recurrent neural network (RNN) is desirable. The composite neural network divides a character string region Tn into a plurality of characters from the leading end to the tail in the character arrangement direction, extracts feature values for each partition by a convolutional neural network, and calculates feature amounts between the leading end and the trailing end. This is a neural network that recognizes a character string by simultaneously capturing a series of recursive neural networks. That is, the composite neural network is a neural network that can recognize a character string as a recognition unit without having to cut out a character string existing in the character string region Tn into individual characters.

  The teacher data for learning given to the composite neural network is font data and an owner code used for the identification character string 20. In this way, the data used for the identification character string 20 of the actual cargo 12 is input to the composite neural network in advance as teacher data, thereby improving the recognition accuracy of the character string using the composite neural network. Will be advantageous. Note that the font data is not limited to one type, but may be various types, rotated at a predetermined angle, or inverted up / down / left / right, and cut out from the actual captured image of the cargo. It is preferable to use letters.

  The first character string recognizing unit 45 receives the data 34 and outputs the first character string 23 recognized by using the composite neural network to the determining unit 43 and the position of the first character string 23 in the second character string recognizing unit 46. Is a functional element to output to The first character string recognizing unit 45 recognizes the first character string 23 based on a known feature that the first character string 23 is a four-digit alphabet and “U” exists as a specific character at the end.

  The second character string recognition unit 46 receives the position of the data 34 and the first character string 23 and outputs the second character string 24 recognized by using the composite neural network to the determination unit 43 and the second character string 24. Is a functional element that outputs the position of the image to the image processing unit 47. The second character string recognition unit 46 recognizes the second character string 24 based on the known feature that the second character string 24 is arranged in the vicinity of the first character string 23 and is a six-digit number.

  The single character recognition unit 42 receives the single character image data 48 (48a, 48b) cut out so that the single character can be discriminated by the image processing unit 47, and outputs the recognized check character 22 to the determination unit 43. It is. The single character recognition unit 42 is a functional element that recognizes the check character 22 using a convolutional neural network with a single character as a recognition unit.

  In addition, when the character string recognizing unit 41 cannot recognize the characteristic character string 21, the single character recognizing unit 42 re-recognizes all or part of the characters constituting the characteristic character string 21 one by one by using a convolutional neural network. It is also a functional element that outputs the re-recognized character string 26 to the character string recognition unit 41 and the determination unit 43.

  An example of the convolutional neural network is LeNet. The teacher data for learning given to the convolutional neural network of the single character recognition unit 42 is font data used for the identification character string 20. Note that the font data is not limited to a single type, as in the case of the teacher data of the complex neural network, but may be a variety of types, rotated at a predetermined angle, or inverted up / down / left / right. It is preferable to use characters cut out from an image obtained by actually capturing cargo.

  The determination unit 43 receives the characteristic character string 21 output from the character string recognition unit 41 and the inspection character 22 output from the single character recognition unit 42 and applies a predetermined calculation formula to the characteristic character string 21. This is a functional element that collates the calculated character 25 and the inspection character 22 calculated in this way. The determination unit 43 is a functional element that collates the calculated character 25 with the check character 22 and outputs the characteristic character string 21 and the check character 22 to the output unit 44 when they match. On the other hand, the determination unit 43 collates the calculated character 25 with the check character 22 and, if they do not match, determines the character constituting the characteristic character string 21 in the single character recognition unit 42 based on the position of the characteristic character string 21. This is a functional element that causes the determination unit 43 to output the re-recognized character string 26 that is re-recognized all or part of the characters one by one as the characteristic character string 21.

  The output unit 44 receives the characteristic character string 21 and the check character 22 output from the determination unit 43, and outputs a character string in which the characteristic character string 21 and the check character 22 are arranged in order from the top as the identification character string 20. This is a functional element for displaying the identification character string 20 output on the display device 31. Further, it is also a functional element that displays “unrecognizable” on the display device 31 when the determination unit 43 instructs the recognition character string 20 to be unrecognizable.

  The image processing unit 47 receives the position of the second character string 24 output from the second character string recognition unit 46 and receives the inspection character image data 48a cut out so that the inspection character 22 can be discriminated. Is a functional element to output to The image processing unit 47 is also a functional element that outputs to the single character recognition unit 42 a plurality of second character string image data 48b in which the character string region Tn in which the second character string 24 exists is divided for each character. The second character string image data 48b has the same number as the number of digits of the second character string 24.

  As illustrated in FIGS. 5 to 9, the method for recognizing the identification character string 20 starts when the data 34 output from the character string region detection system 33 is input, and the identification character string 20 is displayed on the display device 31. Or when “Unrecognizable” is displayed. In addition, the recognition method of the identification character string 20 is a method in which the recognition accuracy is improved as the neural network teacher data is accumulated in the repeated process.

  As illustrated in FIG. 5, when the data 34 is input, the first character string recognizing unit 45 orders a plurality of character string regions Tn existing in the data 34 according to a preset rule P (S110).

  As illustrated in FIG. 10, the data 34 is data in which an area occupied by a character string is detected as a plurality of character string areas Tn (n = 1, 2,..., 25). This is data in which the position of Tn and the occupied area are detected as plane coordinates. The rule P is a rule for setting a priority order for a plurality of character string regions Tn, and is based on the distance between the origin P0 and the character string region Tn with any one of the four corners of the image 14 as the origin P0. It is a rule. In this embodiment, the rule P sets the upper right corner of the image 14 as the origin P0 based on the convention that the identification character string 20 is written in the upper right of the back surface of the cargo 12. The rule P is such that the character string region T3 closest to the origin P0 is the head, the character string region T21 farthest is the tail, and the distance from the origin P0 is close (T3, T2, T4, T1, T8,...・ Ordered to T21).

Next, the first character string recognizing unit 45 recognizes a character string existing in the a-th character string region Tn in the order using the composite neural network (S120). When the character string is written in left horizontal writing, the left end of the character string region Tn may be the beginning and the right end may be the tail, and the number of sections that divide the character string region Tn may be greater than or equal to the number of digits of the first character string 23.

  Next, the first character string recognition unit 45 determines whether or not a preset condition A is satisfied for the recognized character string (S140). Condition A is a condition based on a known feature of the first character string 23. Condition A is preferably satisfied when the recognized character string is composed of a four-digit alphabet and “U” is present as a specific character at the end. Condition A is that the first 3 digits of the recognized character string are B.3. I. More preferably, it is any of the administrator codes registered in C and “U” exists at the end.

  Next, when the condition A is satisfied (S140: YES), the first character string recognizing unit 45 outputs the recognized character string to the determining unit 43 as the first character string 23 and the second character string recognizing unit 46 Data of one character string 23 is output (S150).

  On the other hand, when the condition A is not satisfied (S140: NO), the first character string recognizing unit 45 counts up the order of the order (S160). Next, the first character string recognizing unit 45 determines whether or not the counted up order is larger than the last order ordered by the rule P (S170). If it is determined that the order is greater than the last order, it is determined that the first character string 23 cannot be recognized from the data 34, and “recognition is not possible (ERROR)” is displayed on the display device 31 (S180). On the other hand, if it is determined that the order is equal to or less than the last order, the process returns to step S120, the (a + 1) th character string region Tn in the order is selected, and step S120 and subsequent steps are performed.

  As illustrated in FIG. 6, when the data of the first character string 23 output from the first character string recognition unit 45 is input, the second character string recognition unit 46 performs the character string region Tn according to a preset rule Q. Is selected (S210). The rule Q is a rule for setting priorities for a plurality of character string regions Tn located in the vicinity of the character string region Tn of the first character string 23.

  As illustrated in FIG. 11, the second character string 24 is arranged in the vicinity of the first character string 23. As described above, the second character string 24 is arranged on the right side or below the first character string 23. In addition, the second character string 24 includes a single character string or two character strings divided into two upper and lower columns. Therefore, the rule Q selects a character string region (T2, T3, T4) within a range surrounded by a dotted line in the figure.

  As illustrated in FIG. 6, the second character string recognition unit 46 recognizes a character string existing in the character string region Tn selected using the composite neural network (S220). In this step, the number of sections that divide the character string regions T2, T3, and T4 may be greater than or equal to the number of digits of the second character string 24.

  Next, the second character string recognition unit 46 determines whether or not a preset condition B is established for the recognized character string (S230). Condition B is a condition based on a known feature of the second character string 24. Condition B is one of the conditions that when the recognized character strings are combined, the total number of characters is seven or less and all of them are numbers. Note that the combination of recognized character strings includes the case of one character string. Further, it is preferable that the condition B is satisfied when a six-digit number is formed from the recognized character string. That is, the condition B is that one recognized character string is a six-digit number in one horizontal row, one recognized character string is a seven-digit number in one horizontal row (the last one-digit number is a check character), This is true if one character string consists of six horizontal digits (three digits + three digits, four digits + two digits, etc.), and two recognized character strings consist of two digits in the upper and lower rows. Is more preferable.

In this embodiment, whether or not the condition B is satisfied is determined by selecting a plurality of character string regions Tn in which the second character string 24 is predicted to exist, recognizing the character strings existing therein, Although the determination is based on the combination of columns, the success or failure of condition B may be determined for each character string region Tn in the order of left horizontal writing. For example, the character strings existing in the character string regions T2 to T4 may be recognized in order, and the condition B may be considered to be satisfied when the character string existing in the character string region T2 has six digits. Further, when the character string existing in the character string area T2 is three digits, it is considered that the condition B is satisfied when the character string existing in each of the character string area T3 or the character string area T4 is three digits. May be.

  Next, when the condition B is satisfied (S230: YES), the second character string recognition unit 46 outputs the recognized character string to the determination unit 43 as the second character string 24 and also sends the second character string to the image processing unit 47. The data of the character string 24 is output (S240).

  On the other hand, when the condition B is not satisfied (S230: NO), the second character string recognition unit 46 instructs the character string area detection system 33 to expand the character string area Tn in the character arrangement direction (S250). Next, the second character string recognition unit 46 determines whether or not the expanded character string region Tn has been expanded to the maximum (S260). If it is determined that the character string area Tn has been expanded to the maximum, the second character string 24 cannot be recognized from the data 34, and “recognition not possible (ERROR)” is displayed on the display device 31 (S270). On the other hand, if it is determined that the character string area Tn has not been expanded to the maximum extent, the process returns to step S210, the expanded character string area Tn is selected according to the rule Q, and step S210 and subsequent steps are performed.

  As illustrated in FIG. 7, when the data 34 and the data of the second character string 24 output from the second character string recognition unit 46 are input, the image processing unit 47 converts the data 34 into the data 34 according to a preset rule R. A character string region Tn in which the inspection character 22 is predicted to exist is selected from the plurality of existing character string regions Tn (S310). The rule R is based on the fact that the check character 22 is a single-character number arranged in the vicinity of the tail of the second character string 24, and the characters existing within a predetermined range based on the tail of the second character string 24. This is a rule for selecting the column region T4.

  As illustrated in FIG. 12, the inspection character 22 is arranged near the tail of the second character string 24. As described above, the inspection character 22 is arranged as a single character on the right side or below the tail of the second character string 24. Therefore, the rule R selects the character string region (T3) within the range surrounded by the dotted line in the figure. If a seven-digit number is recognized in the process of success or failure of condition B in step S210, the first six-digit number is the second character string 24, and the last one-digit number is the check character 22. Therefore, the image processing unit 47 may cut out a portion corresponding to the tail end of the target character string region T4.

  As illustrated in FIG. 7, the image processing unit 47 outputs the inspection character image data 48a obtained by performing image processing on the selected character string region Tn to the single character recognition unit 42 (S320).

  Next, the single character recognition unit 42 recognizes a single character existing in the inspection character image data 48a using a convolutional neural network (S330). Next, the single character recognition unit 42 outputs the recognized single character as the check character 22 to the determination unit 43 (S340).

  Next, the determination unit 43 calculates a calculated character 25 by applying a predetermined calculation formula to the first character string 23 and the second character string 24 (S350). The predetermined calculation formula is a calculation formula used when calculating the inspection character 22 from the characteristic character string 21 including the first character string 23 and the second character string 24. Next, the determination unit 43 determines whether or not the calculated character 25 calculated from the first character string 23 and the second character string 24 matches the inspection character 22 recognized by the single character recognition unit 42 (S360).

  As illustrated in FIG. 8, when it is determined that the calculated character 25 matches the inspection character 22 (S <b> 360: YES), the determination unit 43 issues an instruction to the image processing unit 47, and the image processing unit 47 receives the second character string. A plurality of second character string image data 48b subjected to image processing for dividing each character from the character string region Tn where 24 exists is generated (S410). Next, the image processing unit 47 randomly rearranges the plurality of generated second character string image data 48b (S420), and inputs the data to the second character string recognition unit 46 as teacher data (S430).

  Next, the determination unit 43 gives an instruction to the output unit 44. Based on the instruction, the output unit 44 reads the identification string read so that the reading system 40 can read the identification character string 20 on the display device 31. The character string 20 is displayed (S440), and this reading method is completed.

  As illustrated in FIG. 9, when it is determined that the calculated character 25 does not match the inspection character 22 (S <b> 360: NO), the determination unit 43 issues an instruction to the image processing unit 47, and sends the second character to the image processing unit 47. A plurality of second character string image data 48b subjected to image processing for dividing each character from the character string region Tn where the column 24 exists is generated (S510). Next, the image processing unit 47 outputs the generated second character string image data 48b to the second character string recognition unit 46 without changing the order of arrangement (S520).

  Next, the single character recognition unit 42 recognizes single characters existing in the plurality of second character string image data 48b using a convolutional neural network (S530). Next, the single character recognition unit 42 outputs the re-recognition character string 26 to the determination unit 43 (S540).

  Next, the determination unit 43 calculates a recalculated character 28 by applying a predetermined calculation formula to the first character string 23 and the re-recognized character string 26 (S550). Next, the determination unit 43 determines whether or not the calculated recalculated character 28 matches the inspection character 22 (S560). If it is determined that the recalculated character 28 matches the check character 22 (S560: YES), the process proceeds to step S420 in FIG. 8, and this reading method is completed.

  On the other hand, if it is determined that the recalculated character 28 does not match the inspection character 22 (S560: NO), the determination unit 43 issues an instruction to the output unit 44, and the output unit 44 reads the display device 31 based on the instruction. “Recognition is impossible” is displayed so that the system 40 cannot read the identification character string 20 (S570), and this reading method is completed.

  In the above-described reading method, when “unrecognizable” is displayed on the display device 31, the operator visually reads the identification character string 20 described in the cargo 12, and the read character is input to the input device 32 of the terminal. input. Next, the input character string 27 input to the input device 32 is transmitted to the edge device 30, and the edge device 30 can recognize the character string for identification 20 of the cargo 12.

  As described above, in the method of reading the identification character string 20 by the reading system 40, when the identification character string 20 cannot be recognized, teacher data for learning is input in the steps S420 and S430. That is, the accuracy of recognition of the characteristic character string 21 in the characteristic character string recognition unit 41 is increased by repeatedly performing the reading method. Therefore, the process of causing the single character recognition unit 42 to recognize the second character string 24 illustrated in FIG. 9 with time is reduced.

  As described above, in the reading system 40, the character string recognizing unit 41 recognizes the characteristic character string 21 composed of a plurality of characters using a known characteristic as a clue, and the position is specified by the characteristic character string 21 being recognized. The single character recognition unit 42 recognizes the single character as the check character 22.

That is, the image processing step is omitted by causing the character string recognition unit 41 whose recognition unit is a character string to recognize the characteristic character string 21 whose recognition rate is reduced by the image processing step of decomposing the character string into individual characters. Therefore, it is possible to eliminate a decrease in recognition of the characteristic character string 21 due to image processing. On the other hand, in order to eliminate the demerit that the number recognition rate of the character string recognition unit 41 is low, the single character recognition unit 42 whose recognition unit is a single character recognizes the single character check character 22 with high accuracy. Uncertain recognition results of character strings can be eliminated.

  As described above, the reading system 40 pays attention to the fact that the identification character string 20 is composed of the characteristic character string 21 having a known characteristic and the inspection character 22 arranged in the vicinity of the characteristic character string 21. By using two different recognition methods for recognizing the identification character string 20, the disadvantages of the respective recognition methods can be eliminated. Thereby, even when the imaging conditions of the imaging device 13 are poor or when there are a plurality of character strings in the image 14, only the predetermined identification character string 20 can be automatically recognized with high accuracy.

  The reading system 40 is in a state where the recognition rate of the second character string 24 in the second character string recognition unit 46 of the character string recognition unit 41 is low immediately after introduction. In this regard, the recognition rate of the second character string 24 is improved in the process of repeated recognition by supervised learning in which a plurality of second character string image data 48b when recognizable is randomly rearranged and utilized as teacher data. Can do.

  In the reading system 40, the character string recognizing unit 41 can recognize the characteristic character string 21 at a time based on the known characteristic of the identification character string 20, but the characteristic character string 21 is recognized as in the embodiment. It is desirable to classify each known feature and recognize each in turn. Specifically, first, the first character string recognition unit 45 of the character string recognition unit 41 is based on the number of digits of the first character string 23 as a known feature of the identification character string 20, and the character shape and order of the specific character. The first character string 23 is recognized. Next, the second character string recognition unit 46 recognizes the second character string 24 based on the position of the first character string 23 in the image 14 and the number of digits of the second character string 24. In this way, the feature character string 21 is divided into known features and each is recognized in turn, so that each recognition candidate can be greatly narrowed down. This is advantageous for improving the recognition rate, and enables highly accurate automatic recognition. In addition, based on a known feature, the characteristic character string 21 may be divided into three or more, and three or more recognition units capable of recognizing each may be provided.

  In the reading system 40, the character reading method is not particularly limited, but the character recognition unit is converted into a character string using a composite neural network as the character string recognition unit 41, and the recognition unit is set using a convolutional neural network as the single character recognition unit 42. It is desirable to use a single character. A composite neural network in which a convolutional neural network and a recursive neural network are combined is a neural network suitable for prediction and classification of data having meaning in a time series flow. Therefore, such a composite neural network is a neural network suitable for recognizing a combination of a plurality of characters such as the characteristic character string 21. The convolutional neural network is a neural network suitable for detecting features from various parts such as characters and images.

  As described above, the first character string 23 is a combination of the owner code and “U”, and can sufficiently narrow down recognition candidates. On the other hand, the second character string 24 indicates the serial number of the cargo 12, and the recognition candidates cannot be narrowed down. Therefore, the recognition accuracy of the second character string 24 may be lower than the recognition accuracy of the first character string 23. Therefore, when the determination unit 43 determines that the calculated character 25 does not match the check character 22, the single character recognition unit 42 recognizes the second character string 24 character by character, thereby improving the recognition accuracy. .

  The reading system 40 of the above-described embodiment has been described by way of an example of a system that recognizes the identification character string 20 described in the cargo 12 handled at the container terminal 11, but the identification character string 20 has a known feature. If it is a thing, it is also possible to recognize the character string for identification described in those other than the cargo 12. For example, an automobile registration number mark in Japan is exemplified.

In the embodiment described above, the target of the teacher data input in the recognition process has been described by taking the second character string 24 as an example. However, the first character string 23 may be the target of the teacher data in the same manner. Also, the inspection character 22 may be a target of teacher data.

  In the described embodiment, since the first character string 23 includes the characteristic character “U”, the first character string recognition unit 45 recognizes the first character string 23 first, and then the second character string recognition unit 46. However, the present invention is not limited to this configuration. For example, when recognizing an identification character string that includes a characteristic character in the second character string, the second character string may be recognized.

  When “recognition is impossible” is displayed on the display device 31, teacher data may be input based on the input character string 27 input by the input device 32.

20 character string for identification 21 character string 22 inspection character 40 reading system 41 character string recognition unit 42 single character recognition unit 43 determination unit 44 output unit

Claims (6)

An identification character string comprising a characteristic character string having a known characteristic and a single character check character that is arranged in the vicinity of the characteristic character string and that is calculated by applying a predetermined rule to the characteristic character string Is a reading system that reads from an input image,
The character string is a recognition unit from a series of feature amounts between the leading end and the trailing end in the character string area occupied by the character string included in the image, and the character string included in the image is selected from the plurality of character strings included in the image. A character string recognition unit for recognizing the characteristic character string based on a known feature;
Recognizing the inspection character based on the position of the characteristic character string recognized by the character string recognition unit, with a single character as a recognition unit from single character image data cut out so that a single character included in the image can be discriminated. A single character recognition unit to
When the check character matches the calculated character calculated by applying the predetermined rule to the feature character string recognized by the character string recognition unit, the feature character string and the check character are sequentially A reading system comprising: an output unit that recognizes and outputs the arranged character strings as the identification character strings. The characteristic character string includes a first character string including at least one specific character that is configured with a predetermined number of digits of two or more digits and whose shape and order are specified in advance, and is arranged in the vicinity of the first character string. And a second character string composed of two or more digits specified in advance,
The character string recognizing unit recognizes the first character string based on the number of digits of the first character string and the character shape and order of the specific character as the known features, and the image of the recognized first character string. The reading system according to claim 1, wherein the reading system has a function of recognizing the second character string based on the position of the second character string and the number of digits of the second character string. The character string recognizing unit has a function of recognizing the characteristic character string using a composite neural network that combines a convolutional neural network and a recursive neural network,
The reading system according to claim 1, wherein the single character recognition unit has a function of recognizing the check character using a convolutional neural network. A determination unit that determines whether or not the calculated character matches the inspection character and the character string recognition unit has a function of recognizing the characteristic character string using a neural network;
When the determination unit determines that the calculated character does not match the inspection character, the single character recognition unit causes the single character recognition unit to re-recognize all or part of the characters constituting the characteristic character string one by one,
When the determination unit determines that the recalculated character calculated by applying the predetermined rule to the re-recognized character string including the character re-recognized by the single character recognition unit matches the inspection character, The reading according to any one of claims 1 to 3, further comprising a function of randomly sorting image data obtained by dividing the re-recognized character string for each character and inputting the data as teacher data to the character string recognition unit. system. A determination unit that determines whether or not the calculated character matches the inspection character and the character string recognition unit has a function of recognizing the characteristic character string using a neural network;
When the determination unit determines that the calculated character matches the check character, the character string recognition unit randomly determines image data obtained by dividing the character string recognized by the character string recognition unit for each character. 5. The reading system according to claim 1, wherein the reading system has a function of rearranging and inputting the data as teacher data. An identification character string comprising a characteristic character string having a known characteristic and a single character check character that is arranged in the vicinity of the characteristic character string and that is calculated by applying a predetermined rule to the characteristic character string Is a method of reading from an input image,
The character string is a recognition unit from a series of feature amounts between the leading end and the trailing end in the character string area occupied by the character string included in the image, and the character string included in the image is selected from the plurality of character strings included in the image. Recognizing the feature string based on known features;
Examples recognition unit of single characters from the single character image data in which a single character is cut out to be discriminated included in the image, recognizes the check character based on the position of the feature strings began to probe,
Applying said predetermined rule for recognition by said characteristic character string to calculate a calculated character,
Determining whether the calculated character and the inspection character match;
A reading method characterized by recognizing, as the identification character string, the characteristic character string and a character string in which the check characters are arranged in order when it is determined that the calculated character matches the check character.

Images