Steel billet marking character recognition system based on deep learning

Abstract

The present invention relates to a steel billet spray-marking character recognition system based on deep learning, comprising an image acquisition device for acquiring raw image data; an inference server for receiving raw image data and forwarding the raw image data to a recognition module built into the inference server; the inference server is electrically connected to the image acquisition device; and the recognition module is used to perform spray-marking character recognition operations on the raw image data to obtain the slab number. The present invention does not need to reduce the traveling speed of the steel billet to take pictures, thereby improving the efficiency of the production line; adapting to the environment of the collection and recognition location, greatly increasing the applicable setting locations; and recognizing rusted and worn characters, thereby greatly improving the recognition rate.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition, and in particular to a steel billet spray marking character recognition system based on deep learning. Background Art

[0002] Billet is a very common steel shape in steel production and processing enterprises. It is a thick steel plate. In order to distinguish billets, each steel enterprise will mark the billet number on the thick side of the billet.

[0003] On the other hand, the billets are driven by rollers on the assembly line; and modern automatic assembly lines often rely on computer systems to achieve assembly line management; this involves a need to locate each billet for easy management;

[0004] The existing technology uses automatic image recognition to locate the billet. Specifically, a camera is set at a suitable place on the assembly line. When the billet passes by, it is photographed and identified.

[0005] As the planning of modern production lines is very compact, the running speed of the production line is very fast, that is, the speed of the roller is very fast, which drives the billet to move very fast; a negative problem brought about by this is that when the billet number passes by, the photo taken is very blurry, which is not conducive to identification;

[0006] In order to solve this problem, the existing slab number recognition technologies all adopt a solution, that is, where the camera is set, the speed of the roller is forced to be reduced, and the speed of the steel billet is slowed down so that the photo quality meets the recognition requirements;

[0007] On the other hand, the production environment of steel companies is relatively harsh, and not all places are suitable for setting up cameras. For example, places with too strong light are prone to overexposure, while places with too weak light will make the collected photos blurry, which is not conducive to identification. What is even more troubling for steel companies is that the slab number itself will be damaged and lost due to inevitable rust, wear and tear, leaving only a few traces, which is even more difficult to identify. For this type of problem, the existing technology has no technical solution to deal with it. Only a few recognition systems use high-end cameras to slightly improve the recognition problem, but the cost is high and it cannot solve the fundamental problem.

[0008] In summary, the prior art has the following defects:

[0009] 1. In order to meet the photo quality requirements, the billet speed is forced to decrease, which objectively forms a speed bottleneck of the production line. In addition, the speed must be reduced once each set of identification equipment is set up, which restricts the production efficiency of the production line and does not meet the production speed requirements of modern steel companies.

[0010] 2. Since there is no technical solution for collecting and identifying under high light and low light conditions in the prior art, the identification equipment in the prior art has certain requirements for the installation location and cannot be installed anywhere in the production line, which reduces the universality and is not conducive to promotion;

[0011] 3. Since there is no technical solution to the problem of damaged character recognition in the prior art, it is impossible to recognize common partially rusted and worn characters, which reduces the recognition rate and makes the reliability worrying. Summary of the invention

[0012] In response to the above-mentioned problems, the present invention provides a steel billet spray marking character recognition system based on deep learning, the purpose of which is to improve the efficiency of the production line by eliminating the need to reduce the traveling speed of the steel billet for photographing; to adapt to the environment of the collection and identification location and greatly increase the number of applicable setting locations; and to significantly improve the recognition rate of rusted and worn characters.

[0013] To solve the above problems, the technical solution provided by the present invention is:

[0014] The steel billet marking character recognition system based on deep learning includes the following parts:

[0015] An image acquisition device is used to acquire original image data; the original image data is a digital image composed of a plurality of continuous character image data from the steel billet arranged in the acquisition order; each of the character image data is a photo of a spray-marked character; each of the spray-marked characters is a decimal Arabic numeral printed on the side of the steel billet;

[0016] An inference server, used for receiving the original image data and forwarding the original image data to a recognition module built in the inference server; the inference server is electrically connected to the image acquisition device;

[0017] The recognition module is used to perform a spray-mark character recognition operation on the original image data to obtain a slab number; the slab number is a 10-digit decimal string, the value of which is consistent with the information contained in the combination of the identified multiple continuous spray-mark characters in the acquisition order.

[0018] Preferably, the identification module comprises a communication submodule, an identification submodule and a deep learning submodule; wherein:

[0019] a communication submodule, configured to receive the original image data and forward the original image data to the recognition submodule;

[0020] A recognition submodule, used for performing the spray-mark character recognition operation on the original image data to obtain the slab number;

[0021] The deep learning submodule is used to train the recognition AI built into the inference server, thereby improving the recognition accuracy of the steel billet marking character recognition system.

[0022] Preferably, the identification submodule comprises a main identification branch and an auxiliary branch; wherein:

[0023] The auxiliary branch is used to segment the original image data to obtain the character image data, and then store it for use by the deep learning submodule when performing deep learning training, and includes an image segmentation submodule and a character image database; wherein:

[0024] The image segmentation submodule is used to perform segmentation operation on the original image data to obtain the character image data;

[0025] The character image database is used to store the character image data as a deep learning sample for the recognition AI;

[0026] The main recognition branch includes a recognition algorithm for performing the spray-marked character recognition operation on the original image data to obtain the slab number.

[0027] Preferably, the recognition algorithm comprises the following steps:

[0028] S100. Calling the image segmentation submodule, performing the segmentation operation on the original image data in order from left to right, to obtain a plurality of the character image data;

[0029] S200. Classify and recognize the character image data one by one in order from left to right, and perform a probability evaluation operation to obtain the recognized characters and recognition accuracy of each character image data;

[0030] The identification character is a decimal Arabic numeral;

[0031] The recognition accuracy is a percentage value, and its interval is [0.00%, 100.00%];

[0032] S300. Perform a logic determination operation on all the recognized characters identified this time to obtain a logic determination result; the logic determination result includes the character strings "Logical determination passed" and "Logical determination failed";

[0033] S400. Perform the following operations according to the recognition accuracy and the value of the logic determination result:

[0034] If the recognition accuracy rate corresponding to the recognized character is less than the manually preset minimum recognition accuracy rate threshold, or the logic determination result is the string "Logic determination failed", the string "cannot be recognized" is output; then the call to the recognition algorithm is terminated;

[0035] Otherwise, the identification characters are concatenated into a 10-digit decimal value in order from left to right and assigned to the slab number; the slab number is then stored in a suggestion database built into the inference server; the slab number is then output, and the call to the identification algorithm is terminated.

[0036] Preferably, the recognition probability evaluation operation comprises the following steps:

[0037] S210. The recognition AI recognizes the first character image data obtained in this recognition 100 times to obtain 100 recognition points; the recognition points are a binary set of the recognized characters and the recognition accuracy obtained by one recognition;

[0038] S220. According to the 100 identification points in S210, perform the following operations:

[0039] If the 100 identification points do not converge, or do not converge on one identification point, then return to and execute S210 again;

[0040] Otherwise, outputting the recognition point to which the 100 recognition points converge as the recognition character and the recognition accuracy rate of the first character image data;

[0041] S230. Delete the first character image data, and then shift the remaining character image data left by 1 bit in sequence; then execute S240;

[0042] S240. Check whether the current first character image data is empty, and then perform the following operations according to the check result:

[0043] If the first character image data is not empty, return to and execute S210 to S230 again;

[0044] Otherwise, the recognition probability evaluation operation is terminated; and then the recognized characters and the recognition accuracy rate of each character image data are output in sequence according to the recognition order.

[0045] Preferably, the logic determination operation comprises the following steps:

[0046] S310. Check whether the result of the recognition probability evaluation operation is 10 recognized characters, and each recognized character corresponds to 1 recognition accuracy rate, and then perform the following operations according to the check result:

[0047] If the result of the recognition probability evaluation operation is 10 recognized characters, and each recognized character corresponds to 1 recognition accuracy rate, then execute S320;

[0048] Otherwise, the logic determination result is assigned a string "Logic determination failed", and then the logic determination operation is exited;

[0049] S320. Check whether the 10 recognized characters are all Arabic numerals, and then perform the following operations according to the check result:

[0050] If all the 10 recognized characters are Arabic numerals, execute S330;

[0051] Otherwise, the logic determination result is assigned a string "Logic determination failed", and then the logic determination operation is exited;

[0052] S330. Splice the identification characters into a 10-digit decimal value in order from left to right, and assign the slab number; then substitute the slab number into a manually preset slab number verification rule for verification; the slab number verification rule is a manually preset logical expression, and when the slab number meets the slab number verification rule, the verification result is a Boolean value of 1, otherwise it is 0; then perform the following operations according to the verification result:

[0053] If the check result is 1, the logic determination result is assigned a string "Logic determination passed", and then the slab number is output, and then the logic determination operation is exited;

[0054] If the verification result is 0, the logic determination result is assigned a character string "Logic determination failed", and then the logic determination operation is exited.

[0055] Preferably, the image acquisition device comprises an industrial digital camera and a temperature-controlled dustproof protective box; wherein:

[0056] The temperature-controlled dust-proof protective box is a metal box-type container, and a shooting port is provided on the side thereof for the lens of the industrial digital camera to collect the original image data; the horizontal elevation of the shooting port is the same as the horizontal elevation of the side of the slab, and is directly opposite to the side of the slab; a column for supporting function is fixedly installed on the lower end surface of the temperature-controlled dust-proof protective box; the horizontal height of the column can be adjusted;

[0057] The industrial digital camera is detachably mounted inside the temperature-controlled dustproof protective box; the lens of the industrial digital camera faces the shooting port.

[0058] Preferably, the size of the temperature-controlled dustproof protection box is 625mm (H)*325mm (L)*285mm (W).

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1. Since the present invention adopts a deep learning mechanism and uses samples collected in real time for self-training, it can adapt to different billet travel speeds by itself, so there is no need to reduce the billet travel speed for taking pictures, so that identification and collection no longer become the bottleneck of the assembly line, thereby improving the efficiency of the production line;

[0061] 2. Since the present invention adopts a deep learning mechanism, it can perform self-training on samples collected at a specific location, so that different recognition devices can quickly adapt to the environment of the collection and recognition location where they are located, and can greatly increase the number of applicable installation locations, which is very conducive to promotion;

[0062] 3. Since the present invention adopts a deep learning mechanism, it can use the steel billets of a specific enterprise as samples, continuously self-learn, and improve the recognition rate, thereby being able to recognize rusted and worn characters, greatly improving the recognition rate and having extremely high reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic diagram of the logical structure of the identification system according to a specific embodiment of the present invention;

[0064] Figure 2a A photo of a method for setting up an image acquisition device on site according to a specific embodiment of the present invention (left side);

[0065] Figure 2b A photo of a method for setting up an image acquisition device on site according to a specific embodiment of the present invention (right side);

[0066] Figure 3 It is a schematic diagram of a method for setting up an image acquisition device according to a specific embodiment of the present invention.

[0067] Among them: 1. Temperature-controlled dust-proof protective box, 2. Industrial digital camera, 3. Spray-on characters, 4. Steel billet, 5. Roller. DETAILED DESCRIPTION

[0068] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0069] like Figure 1 As shown in the figure, the steel billet marking character recognition system based on deep learning includes the following parts:

[0070] The image acquisition device is used to acquire original image data; the original image data is a digital image composed of multiple continuous character image data from the steel billet 4 arranged in the acquisition order; each character image data is a photo of a spray-marked character 3; each spray-marked character 3 is a decimal Arabic numeral printed on the side; the steel billet 4 is laid on the roller 5 and is transmitted as the roller 5 rotates.

[0071] like Figure 2a and Figure 2b As shown, the image acquisition device includes an industrial digital camera 2 and a temperature-controlled dustproof protection box 1; wherein:

[0072] The temperature-controlled dust-proof protective box 1 is a metal box-type container, and a shooting port is opened on the side for the lens of the industrial digital camera 2 to collect original image data; the horizontal elevation of the shooting port is the same as the horizontal elevation of the side of the slab, and it is facing the side of the slab; a column for support is fixedly installed on the lower end surface of the temperature-controlled dust-proof protective box 1; the horizontal height of the column can be adjusted.

[0073] The industrial digital camera 2 is detachably mounted inside the temperature-controlled dustproof protective box 1; the lens of the industrial digital camera 2 faces the shooting port.

[0074] In this specific embodiment, the size of the temperature-controlled dustproof protection box 1 is 625 mm (H) * 325 mm (L) * 285 mm (W).

[0075] It should be noted that the image acquisition device is installed about 2 meters away from the side of the slab, and the column is adjusted so that the height from the ground is about 1.1 meters. The column is fixed on the ground.

[0076] It should be noted that the difficulty of the image acquisition device lies in the on-site lighting environment and trying not to change the production process.

[0077] This is because the equipment is used in a factory. When the weather is different, the ambient light varies greatly. For example, on a sunny day, the sunlight will make the factory brighter, while on a cloudy day, it will be darker. When the time is different, the light interference is greater, such as turning on the lights at night and not turning on the lights during the day. In addition, the passing of vehicles and the presence of high slabs stacked nearby will also cause changes in the ambient light.

[0078] Based on this situation, the present invention tries not to use light sources, but uses an industrial digital camera 2 with high photosensitivity, and then performs adaptive algorithm adjustment on the changes in illumination from the perspective of software algorithms. Because the light source itself has high heat, and the ambient temperature on site is also high, it is difficult to dissipate heat due to the high temperature on site, which affects the lifespan; in addition, the spray-marked characters 3 are white, which are easy to reflect under strong light, and the ambient light changes on site are complex. When it is too bright to reflect or too dark to be blurred, it is difficult to adjust the illumination in real time. Although dynamic dimming is possible, the overall cost of the system will increase significantly, so research on the configuration of the industrial digital camera 2 and the software algorithm is a relatively better solution.

[0079] like Figure 3 As shown, it is necessary to further explain that, for the production process, although the steel plate will stop for a short time, after on-site investigation, it is found that the stop position is not fixed at a certain point. Sometimes some characters do not enter the field of view, and sometimes some characters are out of the field of view (such as being blocked by the side guard plate of the roller table). If the static collection method is adopted, the roller operator is required to stop the slab at a fixed position so that all characters can enter the field of view completely.

[0080] Therefore, in order to avoid changing the production process as much as possible and provide convenience for the roller operator, the present invention adopts a dynamic acquisition method, that is, using a high-speed and high-sensitivity industrial digital camera 2, combined with a high-speed image processing algorithm, clear character images can still be acquired when the slab is moving.

[0081] Based on the above requirements, in this specific embodiment, the high-sensitivity industrial digital camera 2 adopts a high-speed and high-sensitivity industrial digital camera 2HK-A5100, whose parameters are: resolution 2560*2160, frame rate 17-55fps, dynamic range>83.5dB, and photoelectric efficiency ≥0.55@600nm.

[0082] In order to adapt to HK-A5100, the present invention also builds two modules in the inference server: an image preprocessing module responsible for preprocessing the collected raw image data to form a standard operable format image, and a character image enhancement module for enhancing the contrast between the characters on the image and the surrounding environment; among them, the image preprocessing module has a built-in lighting adaptive algorithm, and the character image enhancement module has a built-in dynamic acquisition algorithm.

[0083] The inference server is used to receive the original image data and forward the original image data to the recognition module built in the inference server; the inference server is connected with the image acquisition device by electrical signals.

[0084] The inference server is placed in the operation room, which is about 10 meters away from the image acquisition device and is connected to the industrial digital camera 2 in the temperature-controlled dustproof protection box 1 by an industrial double-shielded Gigabit network cable.

[0085] The recognition module is used to perform a spray-marked character 3 recognition operation on the original image data to obtain a slab number; the slab number is a 10-digit decimal string, the value of which is consistent with the information contained in the combination of multiple consecutive spray-marked characters 3 to be recognized in the order of acquisition.

[0086] The recognition module includes a communication submodule, a recognition submodule and a deep learning submodule; wherein:

[0087] The communication submodule is used to receive the original image data and forward the original image data to the recognition submodule.

[0088] The recognition submodule is used to perform a spray-marked character 3 recognition operation on the original image data to obtain the slab number.

[0089] In this specific embodiment, the identification submodule includes a main identification branch and an auxiliary branch; wherein:

[0090] The auxiliary branch is used to segment the original image data to obtain character image data, and then store it for use by the deep learning submodule for deep learning training, including an image segmentation submodule and a character image database; wherein:

[0091] The image segmentation submodule is used to perform segmentation operations on the original image data to obtain character image data.

[0092] Character image database, used to store character image data as deep learning samples for recognition AI.

[0093] It should be noted that the auxiliary branch is set up to store and manage character image data and provide sample data for the deep learning submodule parallel to the recognition submodule.

[0094] The main recognition branch includes a recognition algorithm for performing a spray-marked character 3 recognition operation on the original image data to obtain the slab number.

[0095] In this specific embodiment, the recognition algorithm includes the following steps:

[0096] S100. Calling the image segmentation submodule, segmenting the original image data in order from left to right to obtain a plurality of character image data;

[0097] S200. Classify and recognize the character image data one by one in order from left to right, and perform probability evaluation operations to obtain the recognized characters and recognition accuracy of each character image data.

[0098] In this specific embodiment, the recognition probability evaluation operation includes the following steps:

[0099] S210. The first character image data obtained by the recognition AI is recognized 100 times to obtain 100 recognition points; the recognition points are a set of binary sets of recognized characters and recognition accuracy obtained by one recognition;

[0100] S220. According to the 100 identification points in S210, perform the following operations:

[0101] If the 100 identification points do not converge, or do not converge on one identification point, the process returns to and executes S210 again.

[0102] Otherwise, the recognition point to which the 100 recognition points converge is output as the recognition character and recognition accuracy of the first character image data.

[0103] S230. Delete the first character image data, and then shift the remaining character image data left by 1 bit in sequence; then execute S240.

[0104] S240. Check whether the current first character image data is empty, and then perform the following operations according to the check result:

[0105] If the current first character image data is not empty, the process returns and executes S210 to S230 again.

[0106] Otherwise, the current recognition probability evaluation operation is terminated; and then the recognized characters and recognition accuracy of each character image data are output in sequence according to the recognition order.

[0107] Recognizes a character as a decimal numeral.

[0108] The recognition accuracy is a percentage value, and its range is [0.00%, 100.00%].

[0109] S300. Perform a logic determination operation on all the recognized characters identified this time to obtain a logic determination result; the logic determination result includes character strings "Logical determination passed" and "Logical determination failed".

[0110] In this specific embodiment, the logic determination operation includes the following steps:

[0111] S310. Check whether the result of the recognition probability evaluation operation is 10 recognized characters, and each recognized character corresponds to 1 recognition accuracy rate, and then perform the following operations according to the check result:

[0112] If the result of the recognition probability evaluation operation is 10 recognized characters, and each recognized character corresponds to 1 recognition accuracy rate, S320 is executed.

[0113] Otherwise, the logic judgment result is assigned the string "Logic judgment failed", and then the logic judgment operation is exited.

[0114] S320. Check whether all 10 recognized characters are Arabic numerals, and then perform the following operations based on the check result:

[0115] If the 10 recognized characters are all Arabic numerals, S330 is executed.

[0116] Otherwise, the logic judgment result is assigned the string "Logic judgment failed", and then the logic judgment operation is exited.

[0117] S330. Splice the identified characters into a 10-digit decimal value in order from left to right, and assign the slab number; then substitute the slab number into the manually preset slab number verification rule for verification; the slab number verification rule is a manually preset logical expression, when the slab number meets the slab number verification rule, the verification result is a Boolean value of 1, otherwise it is 0; then perform the following operations according to the verification result:

[0118] If the verification result is 1, the logic judgment result is assigned the string "Logic judgment passed", and then the slab number is output, and then the logic judgment operation is exited.

[0119] If the verification result is 0, the logic judgment result is assigned the string "Logic judgment failed", and then the logic judgment operation is exited.

[0120] It should be noted that the physical meaning of the slab number verification rule is a series of artificially preset logical rules; in the simplest case, the previous and next slab numbers have certain association rules. If the slab number recognized once does not conform to this association rule, it can be determined that the recognition is wrong; for example, the previous and next slab numbers are in an arithmetic progression of natural numbers, and the arithmetic difference value is 1. When the previous recognition results have been... 1738947231, 1738947232, 1738947233, and suddenly 1738947230 appears, there may be a problem, indicating that the recognition may be wrong this time, and the result is unreliable and needs to be discarded.

[0121] S400. Perform the following operations according to the recognition accuracy and the value of the logic determination result:

[0122] If there is a recognition character whose recognition accuracy is less than the manually preset minimum recognition accuracy threshold, or the logic judgment result is the string "logical judgment failed", the string "cannot be recognized" is output; then the call to the recognition algorithm is terminated.

[0123] Otherwise, the recognized characters are concatenated into a 10-digit decimal value from left to right and assigned a slab number; the slab number is then stored in a suggestion database built into the inference server; the slab number is then output, and the call to the recognition algorithm is terminated.

[0124] It should be noted that the present invention tests the recognition results from two aspects: recognition accuracy and logical judgment results, which greatly reduces the recognition error rate. This is obviously improved significantly compared with the existing technology in practical applications. At present, the accuracy rate is close to 99%.

[0125] It should be further explained that, although the recognition rate of the present invention is very high, the slab number may still be damaged during transportation, warehousing, etc. In order not to affect the work efficiency and the normal process of the production line, the present invention also sets a manual modification interface in the recognition submodule and makes corresponding adaptations on the UI. In this way, if a recognition error is found, but there is no problem with the steel billet 4 itself, manual adjustment can be made to facilitate interactive modification without stopping the production line for problems with individual slab number characters.

[0126] It should be further explained that the recognition results will eventually be stored in the suggestion database for future query.

[0127] The deep learning submodule is used to train the recognition AI built into the inference server, thereby improving the recognition accuracy of the billet marking character recognition system.

[0128] It should be noted that the biggest difference between the present invention and the prior art based on deep learning is that as the usage time increases, the collected character image samples increase, and through data learning, the system performance can be improved, that is, the recognition accuracy is higher.

[0129] Specifically, the deep learning submodule of the present invention will collect a certain amount of character image data samples after a period of use, and start deep learning training through manual annotation. Based on a certain amount of samples, training the recognition AI to perform data learning will increase the recognition rate.

[0130] It should be further explained that since the first learning and training effect is greatly improved, the second learning and training effect is relatively insignificant, but it is necessary.

[0131] In order to verify the effectiveness of the AI ​​recognition in the present invention, the inventors also conducted a double-blind experiment; that is, the parameters of the untrained system were fixed, two sets were copied, and two independent experiments were conducted in the same environment without human intervention. Table 1 shows the results of the two experiments:

[0132] Table 1. Double-blind experiment results data table

[0133] Initial recognition rate One-time learning recognition rate Secondary learning recognition rate Three-time learning recognition rate Double-blind group A 78.66％ 90.84％ 92.87％ 94.54％ Double-blind group B 78.66％ 90.15％ 93.52％ 96.75％

[0134] On the other hand, the inventors further applied the system to conduct multiple tests in different environments, and verified the recognition reliability of the system; no matter what the initial recognition rate is, after more than 6 times of learning, the recognition rate can approach 99%. The reason why it cannot exceed 99% in the end is that various extreme situations cannot be avoided, such as characters being rusted and unable to be recognized by people, and this statistical noise is tolerable.

[0135] It should be further explained that the software part of the present invention is written in C++ language, which organically integrates the conventional deep learning algorithm based on Python language and the underlying algorithm of industrial digital camera 2 based on C++ language SDK development kit.

[0136] Deep learning algorithms are generally implemented using Python, but software written in Python cannot be used in industrial production for three reasons: 1. Software for industrial production needs to run for a long time and needs to be maintenance-free, while programs written in Python are not very stable; 2. Users need to have a basic understanding of Python, such as inputting operating instructions, etc.; 3. Industrial digital cameras 2 involve the underlying hardware, which is basically written in C and C++. Although C++ and Python programming have interfaces, they have low operating efficiency and are prone to various problems. The software part of the present invention uses the C++ source code of Baidu Feijiang deep learning framework to modify the underlying recognition algorithm and unify it into C++ language development.

[0137] In the above detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are clearly stated in each claim. On the contrary, as reflected in the appended claims, the invention is in a state of having less than all the features of the disclosed individual embodiments. Therefore, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0138] The disclosed embodiments are described above to enable any person skilled in the art to implement or use the present invention. Various modifications of these embodiments are obvious to those skilled in the art, and the general principles defined herein may also be applied to other embodiments without departing from the spirit and scope of the present disclosure. Therefore, the present disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0139] The above description includes examples of one or more embodiments. Of course, it is not possible to describe all possible combinations of components or methods for the purpose of describing the above embodiments, but it should be recognized by those skilled in the art that the various embodiments may be further combined and arranged. Therefore, the embodiments described herein are intended to cover all such changes, modifications and variations that fall within the scope of protection of the appended claims. In addition, with respect to the term "comprising" used in the specification or claims, the word is covered in a manner similar to the term "including", just as "including," is explained as a transitional word in the claims. In addition, any term "or" used in the specification of the claims is intended to mean "non-exclusive or".

[0140] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A steel billet marking character recognition system based on deep learning, characterized by: Contains the following sections: An image acquisition device is used to acquire original image data; the original image data is a digital image composed of a plurality of continuous character image data from a steel billet (4) arranged in an acquisition order; each of the character image data is a photo of a spray-marked character (3); each of the spray-marked characters (3) is a decimal Arabic numeral printed on the side of the steel billet (4); An inference server, configured to receive the original image data and forward the original image data to a recognition module built into the inference server; The inference server is electrically connected to the image acquisition device; A recognition module is used to perform a spray-marked character (3) recognition operation on the original image data to obtain a slab number; the slab number is a 10-digit decimal string, the value of which is consistent with the information contained in the combination of the identified multiple continuous spray-marked characters (3) in the order of acquisition; The identification module includes a communication submodule, an identification submodule and a deep learning submodule; wherein: a communication submodule, configured to receive the original image data and forward the original image data to the recognition submodule; A recognition submodule, used for performing the spray-marked character (3) recognition operation on the original image data to obtain the slab number; A deep learning submodule, used for training the recognition AI built into the inference server, thereby improving the recognition accuracy of the steel billet (4) spray-marked character (3) recognition system; The identification submodule comprises a main identification branch and an auxiliary branch; wherein: The auxiliary branch is used to segment the original image data to obtain the character image data, and then store it for use by the deep learning submodule when performing deep learning training, and includes an image segmentation submodule and a character image database; wherein: The image segmentation submodule is used to perform segmentation operation on the original image data to obtain the character image data; The character image database is used to store the character image data as a deep learning sample for the recognition AI; The main recognition branch includes a recognition algorithm for performing the spray-marked character (3) recognition operation on the original image data to obtain the slab number; The recognition algorithm comprises the following steps: S100. Calling the image segmentation submodule, performing the segmentation operation on the original image data in order from left to right, to obtain a plurality of the character image data; S200. Classify and recognize the character image data one by one in order from left to right, and perform a probability evaluation operation to obtain the recognized characters and recognition accuracy of each character image data; The identification character is a decimal Arabic numeral; The recognition accuracy is a percentage value, and its interval is [0.00%, 100.00%]; S300. Perform a logic determination operation on all the recognized characters identified this time to obtain a logic determination result; the logic determination result includes the character string "Logical determination passed" and "Logical determination failed"; S400. Perform the following operations according to the recognition accuracy and the value of the logic determination result: If the recognition accuracy rate corresponding to the recognized character is less than the manually preset minimum recognition accuracy rate threshold, or the logic determination result is the string "logical determination failed", the string "cannot be recognized" is output; then the call to the recognition algorithm is terminated; Otherwise, the identification characters are concatenated into a 10-digit decimal value in order from left to right, and assigned to the slab number; the slab number is then stored in a suggestion database built into the inference server; the slab number is then output, and the call to the identification algorithm is terminated; The recognition probability evaluation operation comprises the following steps: S210. The recognition AI recognizes the first character image data obtained in this recognition 100 times to obtain 100 recognition points; the recognition points are a binary set of the recognized characters and the recognition accuracy obtained by one recognition; S220. According to the 100 identification points in S210, perform the following operations: If the 100 identification points do not converge, or do not converge on one identification point, then return to and execute S210 again; Otherwise, outputting the recognition point to which the 100 recognition points converge as the recognition character and the recognition accuracy rate of the first character image data; S230. Delete the first character image data, and then shift the remaining character image data left by 1 bit in sequence; then execute S240; S240. Check whether the current first character image data is empty, and then perform the following operations according to the check result: If the first character image data is not empty, return to and execute S210 to S230 again; Otherwise, the recognition probability evaluation operation is terminated; and then the recognized characters and the recognition accuracy rate of each character image data are outputted in sequence according to the recognition order; The logic determination operation comprises the following steps: S310. Check whether the result of the recognition probability evaluation operation is 10 recognized characters, and each recognized character corresponds to 1 recognition accuracy rate, and then perform the following operations according to the check result: If the result of the recognition probability evaluation operation is 10 recognized characters, and each recognized character corresponds to 1 recognition accuracy rate, then execute S320; Otherwise, the logic determination result is assigned a string "Logic determination failed", and then the logic determination operation is exited; S320. Check whether the 10 recognized characters are all Arabic numerals, and then perform the following operations according to the check result: If all the 10 recognized characters are Arabic numerals, execute S330; Otherwise, the logic determination result is assigned a string "Logic determination failed", and then the logic determination operation is exited; S330. Splice the identification characters into a 10-digit decimal value in order from left to right, and assign the slab number; then substitute the slab number into a manually preset slab number verification rule for verification; the slab number verification rule is a manually preset logical expression, and when the slab number meets the slab number verification rule, the verification result is a Boolean value of 1, otherwise it is 0; then perform the following operations according to the verification result: If the check result is 1, the logic determination result is assigned a string "Logic determination passed", and then the slab number is output, and then the logic determination operation is exited; If the verification result is 0, the logic determination result is assigned a string "Logic determination failed", and then the logic determination operation is exited.

2. The steel billet marking character recognition system based on deep learning according to claim 1 is characterized in that: The image acquisition device comprises an industrial digital camera (2) and a temperature-controlled dustproof protection box (1); wherein: The temperature-controlled dust-proof protective box (1) is a metal box-type container, and a shooting port is provided on the side thereof for the lens of the industrial digital camera (2) to collect the original image data; the shooting port has the same horizontal elevation as the side of the slab and faces the side of the slab; a column for supporting is fixedly mounted on the lower end surface of the temperature-controlled dust-proof protective box (1); the horizontal height of the column can be adjusted; The industrial digital camera (2) is detachably mounted inside the temperature-controlled dustproof protective box (1); the lens of the industrial digital camera (2) faces the shooting port.

3. The steel billet marking character recognition system based on deep learning according to claim 2 is characterized in that: The dimensions of the temperature-controlled dust-proof protection box (1) are 625 mm (H)*325 mm (L)*285 mm (W).

Images