Character code recognition method, readable storage medium and computer equipment

Abstract

The invention provides a character code recognition method, a readable storage medium and computer equipment, and relates to the technical field of visual recognition of characters on a workpiece plate. The character code recognition method comprises the following steps: respectively training a detection model and a recognition model by utilizing training character code data through a transfer learning strategy; respectively converting the trained detection model and the trained recognition model into a corresponding detection reasoning model and a corresponding recognition reasoning model; collecting a character coding image at the workpiece; using the detection reasoning model to obtain encoding position plane attitude information of a plurality of acquired character strings in the character encoding image; using the identification reasoning model to obtain information identification results of the plurality of acquired character strings based on the encoding position plane attitude information of the plurality of acquired character strings; and comparing the information identification results of the plurality of collected character strings with all standard character strings in a standard character library to obtain an optimal comparison result of each collected character string.

Description

Technical Field

[0001] The present invention relates to the technical field of visual recognition of characters on a workpiece plate, and more specifically, to a character coding recognition method, a readable storage medium, and a computer device. Background Art

[0002] In fields such as shipyards that require large-scale use of automatic visual recognition, automatic welding, and cutting, in order to improve the automation level, reduce the labor intensity of personnel, and improve the operation efficiency, labels are pasted on large workpieces, such as plate-shaped workpieces in shipyards. The labels record relevant information of the workpieces, such as dimensions, uses, materials, etc., or only record corresponding digital codes, such as barcodes, etc. Then, by collecting images of the labels, all the information of the workpieces can be automatically recognized, so as to facilitate a detailed understanding of all the characteristics of specific workpieces during subsequent automatic welding and cutting.

[0003] However, due to the large volume of such workpieces and poor storage environments, and in addition, the environment in factory areas such as shipyards is also poor, scratches, oil stains, footprints, and even unclear printing and ghosting often occur at the character positions of the labels.

[0004] This causes inaccurate recognition by the visual recognition system, and there are often cases of incorrect recognition or failure to recognize, which cannot effectively cooperate with subsequent automatic welding and cutting. Summary of the Invention

[0005] To solve the above problems, the present invention provides a character coding recognition method, including:

[0006] Training a detection model and an identification model respectively using training character coding data through a transfer learning strategy;

[0007] Converting the trained detection model and the identification model into corresponding detection inference models and identification inference models respectively;

[0008] Collecting a character coding image at the workpiece;

[0009] Using the detection inference model to obtain the coding position plane attitude information of multiple collected strings in the character coding image;

[0010] Using the identification inference model to obtain information recognition results of multiple collected strings based on the coding position plane attitude information of multiple collected strings;

[0011] Comparing the information recognition results of multiple collected strings with all standard strings in a standard character library to obtain the optimal comparison result of each collected string.

[0012] Optionally, before training the detection model and the recognition model using the training character encoding data through a transfer learning strategy, it further includes:

[0013] Collecting original character encoding data;

[0014] Labeling the original character encoding data;

[0015] Performing enhancement and amplification processing on the labeled original character encoding data to obtain the training character encoding data;

[0016] Among them, the enhancement and amplification processing includes at least one of affine transformation, scaling, brightness increase and decrease, and dilation and erosion.

[0017] Optionally, training the detection model and the recognition model using the training character encoding data through a transfer learning strategy includes:

[0018] Adjusting parameters, adopting a transfer learning strategy, and combining the training character encoding data to train the DBNet character detection model in PaddleOCR;

[0019] Adjusting parameters, adopting a transfer learning strategy, and combining the training character encoding data to train the SVTR-PPLCNetV3 character recognition model in PaddleOCR;

[0020] Among them, the detection model is the DBNet character detection model, and the recognition model is the SVTR-PPLCNetV3 character recognition model.

[0021] Optionally, after converting the trained detection model and the recognition model into corresponding detection inference models and recognition inference models respectively, it includes:

[0022] Deploying the detection model and the recognition model in X64-CPU using the Paddle deep learning inference library.

[0023] Optionally, using the detection inference model to obtain the encoding position plane pose information of multiple collected strings in the character encoding image includes:

[0024] Using the detection inference model to detect the four-point coordinates of the rectangles where multiple collected strings are located;

[0025] Processing the four-point coordinates of the rectangles where multiple collected strings are located to obtain corresponding areas of multiple detected character regions;

[0026] Setting external parameters to filter out false detections for the areas of multiple detected character regions;

[0027] Affinely transform the longest string among the multiple collected strings into a horizontal image according to the angle between the longest string and the horizontal direction;

[0028] Perform a secondary detection on the horizontal image to obtain the coordinate information of the horizontal image;

[0029] Perform an inverse transformation check on the coordinate information of the horizontal image.

[0030] Optionally, using the recognition inference model, based on the encoded position plane pose information of the multiple collected strings, obtaining the information recognition results of the multiple collected strings includes:

[0031] According to the four-point coordinates of the rectangle where the multiple collected strings are located, and the corresponding areas of the multiple detected character regions, segment the corresponding region images where the multiple collected strings are located from the character encoding image;

[0032] Use the recognition inference model to score the original images of the corresponding region images where the multiple collected strings are located and their rotated images rotated 180 degrees respectively;

[0033] Perform histogram equalization enhancement on the original images and / or the rotated images with scores lower than the set score, and score again;

[0034] Select the higher scorers among the multiple original images and the multiple rotated images, and those higher than the set score, as the information recognition results of the multiple collected strings.

[0035] Optionally, before comparing the information recognition results of the multiple collected strings with all the standard strings in the standard character library to obtain the optimal comparison result of each collected string, include:

[0036] Design a string merging algorithm according to the character distribution characteristics of the shipyard workpieces;

[0037] Use the string merging algorithm and set an external parameter distance threshold to merge the multiple collected strings;

[0038] According to the information recognition results of the multiple collected strings, and the angle between the longest string among the multiple collected strings and the horizontal direction, calculate and obtain the encoded position plane pose information of the merged multiple collected strings.

[0039] Optionally, after comparing the information recognition results of the multiple collected strings with all the standard strings in the standard character library to obtain the optimal comparison result of each collected string, include:

[0040] Output the string encoding information of each of the collected strings based on the optimal alignment result of each of the collected strings.

[0041] In addition, the present invention also provides a readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the character encoding recognition method described above is implemented.

[0042] In addition, the present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the character encoding recognition method described above is implemented.

[0043] The technical effects of the present invention at least include:

[0044] On the one hand, by using the transfer learning strategy and cooperating with the already trained training character encoding data, the detection model and the recognition model are respectively trained. In this way, for the special situation of few shipyard labels, the data collection of label characters under special shipyard working conditions is greatly reduced. Only need to first train the pre-trained model with relevant common data, such as common character data in advertisements, printing, etc. After the training is completed, then through the transfer learning strategy, use the training character encoding data to train the detection model and the recognition model respectively, so as to ensure that the two models reach a very high recognition accuracy after training. After being converted into corresponding inference models, the detection and recognition efficiency of the detection inference model and the recognition inference model is improved. In this way, both the detection accuracy of the two models is ensured and their detection efficiency is improved.

[0045] On the other hand, use the detection inference model to obtain the encoding position plane attitude information of the collected strings, and use the recognition inference model to obtain the information recognition results of the multiple collected strings to ensure the accuracy of character recognition. Then compare with all the standard strings in the standard character library, and use all the standard strings in the standard character library to verify the detection results to prevent interference from affecting the identification of the character encoding image at the workpiece due to scratches, oil stains, footprints, or even unclear printing and ghosting. And based on this, the detection and recognition results are obtained, greatly improving the accuracy of character recognition on the label.

[0046] In this way, through the mutual combination of the above two aspects, for the situation where there are often scratches, oil stains, footprints, or even unclear printing and ghosting at the characters of the shipyard workpiece labels, the recognition accuracy of the visual recognition system in this situation is effectively improved. It realizes effective cooperation with subsequent automatic welding and cutting. Description of the Drawings

[0047] Figure 1 It is a main schematic flowchart of the character encoding recognition method of the specific embodiment of the present invention;

[0048] Figure 2Schematic flowchart for obtaining training character encoding data in the specific embodiment of the present invention;

[0049] Figure 3 Schematic flowchart for obtaining the encoding position plane attitude information of multiple acquisition strings in the character encoding image in the specific embodiment of the present invention;

[0050] Figure 4 Schematic flowchart for obtaining the information recognition results of multiple said acquisition strings in the specific embodiment of the present invention;

[0051] Figure 5 Schematic flowchart for obtaining the encoding position plane attitude information of multiple merged said acquisition strings in the specific embodiment of the present invention;

[0052] Figure 6 All schematic flowcharts of the character encoding recognition method in the specific embodiment of the present invention. Specific Embodiment

[0053] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the accompanying drawings. Many specific details are set forth in the following description to fully understand the embodiments of the present invention. It should be understood that the specific embodiments described herein are merely used to explain the present invention and are not used to limit the present invention. The embodiments of the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0054] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0055] See Figures 1 to 6 , this embodiment provides a character encoding recognition method, including:

[0056] By using a transfer learning strategy, training a detection model and an identification model respectively with training character encoding data;

[0057] Converting the trained detection model and the identification model into corresponding detection inference models and identification inference models respectively;

[0058] Collecting the character encoding image at the workpiece;

[0059] Using the detection inference model, obtain the encoded position plane attitude information of multiple collected strings in the character encoding image;

[0060] Using the recognition inference model, based on the encoded position plane attitude information of multiple collected strings, obtain the information recognition results of multiple collected strings;

[0061] Compare the information recognition results of multiple collected strings with all standard strings in the standard character library to obtain the optimal comparison result for each collected string.

[0062] In this embodiment, on the one hand, using the transfer learning strategy and cooperating with the already trained training character encoding data, train the detection model and the recognition model respectively. In this way, for the special situation of few shipyard labels, the data collection of label characters under special shipyard working conditions is greatly reduced. Only need to first train the pre-trained model with relevant common data, such as advertising, printing and other common character data. After the training is completed, then use the transfer learning strategy to train the detection model and the recognition model respectively with the training character encoding data, so as to ensure that the two models reach a very high recognition accuracy after training. After converting to the corresponding inference models, the detection and recognition efficiency of the detection inference model and the recognition inference model is improved. In this way, both the detection accuracy of the two models is ensured and their detection efficiency is improved.

[0063] On the other hand, use the detection inference model to obtain the encoded position plane attitude information of the collected strings, and the recognition inference model to obtain the information recognition results of multiple collected strings, to ensure the accuracy of character recognition, and then compare with all standard strings in the standard character library, and use all standard strings in the standard character library to verify the detection results, preventing interference from scratches, oil stains, footprints, and even unclear printing and double images that affect the recognition of the character encoding image at the workpiece. And based on this, obtain the detection and recognition results, greatly improving the accuracy of character recognition on the label.

[0064] In this way, through the combination of the above two aspects, for the situation where there are often scratches, oil stains, footprints, and even unclear printing and double images at the character part of the shipyard workpiece label, the recognition accuracy of the visual recognition system in this situation is effectively improved. Realize effective cooperation with subsequent automatic welding and cutting.

[0065] See Figures 1 to 6 , further, before training the detection model and the recognition model respectively with the training character encoding data through the transfer learning strategy, it also includes:

[0066] Collect the original character encoding data;

[0067] Label the original character encoding data;

[0068] Perform enhancement and amplification processing on the labeled original character encoding data to obtain the training character encoding data;

[0069] Among them, the enhancement and amplification processing includes at least one of affine transformation, scaling, brightness increase and decrease, and dilation and erosion.

[0070] The original character encoding data here can be label image data of real scenes with scratches, oil stains, footprints, unclear and overlapping prints, etc. collected from multiple workpieces in a shipyard.

[0071] Annotate the original character encoding data. Doing so enables the computer to better understand this data, thereby performing more accurate learning and processing. In this way, specific marks are added to the data represented by character encoding, making them easier to understand and process. Thus, the training efficiency of the subsequent model is improved.

[0072] See Figures 1 to 6 , further, through the transfer learning strategy, using the training character encoding data to train the detection model and the recognition model respectively includes:

[0073] Adjust the parameters, adopt the transfer learning strategy, and combine the training character encoding data to train the DBNet character detection model in PaddleOCR;

[0074] Adjust the parameters, adopt the transfer learning strategy, and combine the training character encoding data to train the SVTR-PPLCNetV3 character recognition model in PaddleOCR;

[0075] Among them, the detection model is the DBNet character detection model, and the recognition model is the SVTR-PPLCNetV3 character recognition model.

[0076] With the multi-language support of PaddleOCR, PaddleOCR can not only recognize Chinese, but also recognize texts in multiple languages. Moreover, PaddleOCR has a very high recognition accuracy and strong recognition ability. Even if the text is blurred or the background is complex, it can still accurately recognize the text. It is especially suitable for the situation where the label characters of shipyard workpieces have scratches, oil stains, footprints, or even unclear printing and ghosting. In addition, PaddleOCR has a powerful detection ability. PaddleOCR can detect the position of the text in the image. Whether the text is horizontal, vertical or oblique, it can find it, and it supports text in different directions: some texts may be horizontal, some may be vertical, and PaddleOCR can recognize texts in various directions. Especially for the labels at the workpieces in the shipyard, they are often pasted randomly, and the characters are pasted horizontally, vertically or obliquely, which increases the detection difficulty of visual recognition. And using PaddleOCR can exactly overcome this problem. Moreover, PaddleOCR has a fast processing speed and can recognize text in real time. The above advantages of PaddleOCR are mainly reflected in the DBNet character detection model, thus fully improving the detection ability, detection real-time performance, recognition accuracy and recognized character types of the detection model.

[0077] The SVTR-PPLCNetV3 character recognition model has the advantages of high efficiency and multi-language adaptation. Moreover, SVTR introduces local and global hybrid blocks, which are used to extract character component features and inter-character dependencies respectively. Together with multi-scale features, they form a multi-granularity feature description. And SVTR has achieved state-of-the-art performance in recognizing English and Chinese scene texts and has the characteristic of high accuracy. SVTR-T has the characteristic of being lightweight. It is an effective, smaller and faster model, which is very suitable for use in scenarios with limited resources. It has 6.03M parameters, and on an NVIDIA 1080Ti GPU, it takes an average of 4.5 milliseconds to process each image text. And the recognition accuracy is higher. Compared with PP-OCRv2, PP-OCRv3 has an over 5% improvement in the recognition accuracy in Chinese scene and an 11% improvement in English and digital scenes. SVTR-T can learn the features of a character by regarding it as a whole and successfully capture the dependencies between different characters, thus accurately recognizing the line-wrapped strings on the label.

[0078] In this way, using shipyard data, the DBNet character detection model and SVTR-PPLCNetV3 character recognition model of PaddleOCR are trained through the transfer learning strategy. It makes corresponding preparations for the accuracy and adaptability of character detection in subsequent various complex situations.

[0079] See Figures 1 to 6, further, after converting the trained detection model and the recognition model into corresponding detection inference models and recognition inference models respectively, it includes:

[0080] Deploy the detection model and the recognition model in the X64-CPU using the Paddle deep learning inference library.

[0081] See Figures 1 to 6 , further, using the detection inference model, obtaining the encoding position plane pose information of multiple collected strings in the character encoding image includes:

[0082] Use the detection inference model to detect the four-point coordinates of the rectangular boxes where multiple collected strings are located;

[0083] Process the four-point coordinates of the rectangular boxes where multiple collected strings are located to obtain corresponding areas of multiple detected character regions;

[0084] Set external parameters to filter out false detections for the areas of multiple detected character regions;

[0085] According to the angle between the longest string among multiple collected strings and the horizontal direction, affine-transform the longest string among multiple collected strings into a horizontal image;

[0086] Perform secondary detection on the horizontal image to obtain the coordinate information of the horizontal image;

[0087] Perform inverse transformation verification on the coordinate information of the horizontal image.

[0088] After performing secondary detection and transformation, transform the transformed coordinate information back to its original state to ensure its accurate position on the original picture.

[0089] See Figures 1 to 6 , further, using the recognition inference model, obtaining the information recognition results of multiple collected strings based on the encoding position plane pose information of multiple collected strings includes:

[0090] According to the four-point coordinates of the rectangular boxes where multiple collected strings are located and the corresponding areas of multiple detected character regions, segment the corresponding region images where multiple collected strings are located from the character encoding image;

[0091] Use the recognition inference model to score the original image and the rotated image rotated 180 degrees of the corresponding region images where multiple collected strings are located respectively;

[0092] Histogram equalization enhancement is performed on the original image and / or the rotated image with a score lower than the set score, and the scoring is performed again; realizing the secondary recognition of characters with low scores, and adopting histogram equalization enhancement during the secondary recognition process, improving the accuracy of the secondary recognition. Reducing the situation of equal scores or lower scores after secondary recognition using the same method.

[0093] Select the high scorers among multiple original images and multiple rotated images, as well as those higher than the set score, as the information recognition results of multiple collected strings. Improve the accuracy of recognition and make the screening process more reasonable.

[0094] By scoring the original image of the corresponding area where multiple collected strings are located and the rotated image after rotating it 180 degrees respectively, the accuracy of recognizing upside-down characters is improved.

[0095] See Figures 1 to 6 , further, before obtaining the optimal comparison result of each collected string by comparing the information recognition results of multiple collected strings with all standard strings in the standard character library, it includes:

[0096] Design a string merging algorithm according to the character distribution characteristics of the shipyard workpieces;

[0097] Use the string merging algorithm and set an external parameter distance threshold to merge multiple collected strings;

[0098] According to the information recognition results of multiple collected strings and the included angle between the longest string among multiple collected strings and the horizontal direction, calculate the coded position plane attitude information of the merged multiple collected strings.

[0099] The workpieces here can be workpieces with regular shapes, such as ultra-long plates, triangular plates, rectangular plates and other workpieces with various different shapes.

[0100] Considering the current automated production requirements of the shipyard, during the automatic lifting process of workpieces, especially during the lifting process of large and ultra-long workpieces, the end effector needs to be set at a preset angle in advance and then adsorbed and grabbed. Therefore, the rotation angle of the end effector before hoisting needs to be verified in advance.

[0101] Under normal circumstances, an image acquisition device, such as a ccd camera or other camera, is installed on an automatic traveling device, such as a gantry, a robotic arm, etc. These automatic traveling devices all move in a straight line along a set direction, and during this process, the image acquisition device is used to collect labels. And the labels are usually pasted along the length direction of the workpiece. That is to say, the length direction of the normal string is parallel or perpendicular to the long side of the workpiece.

[0102] The calculated encoded position plane attitude information of the multiple collected strings after merging is more accurate than the encoded position plane attitude information of each string before merging, making the coordinates of the label and the rotation angle relative to the workpiece obtained based on the encoded position plane attitude information more precise. Then, based on this data and the linear motion information of the image acquisition device along the set direction, the inclination angle and position coordinates of the workpiece relative to the straight line in the set direction are obtained. Then, based on this, it provides effective support and multiple verifications for the movement of the end effector, especially the movement of the end effector with a huge volume, and the relevant parameters of changing its lifting axis and rotation axis before grasping.

[0103] For example, verify whether the position of the end effector movement end point is correct by whether the character coordinates are close to the workpiece coordinates; or use the rotation angle information of the merged string relative to the horizontal direction to verify whether the inclination angle of the workpiece recognized by vision meets the requirements.

[0104] Realize using the encoded position plane attitude information of the multiple collected strings after merging to ensure the movement accuracy of grasping devices such as end effectors.

[0105] See Figures 1 to 6 , further, after comparing the information recognition results of the multiple collected strings with all the standard strings in the standard character library and obtaining the optimal comparison result for each collected string, it includes:

[0106] Based on the optimal comparison result of each collected string, output the string encoding information of each collected string.

[0107] Then, based on the string encoding information of each collected string, calculate the four-point coordinates of the rectangle where the corresponding collected string is located and the character encoding position plane attitude information, and use them as the recognition encoding result.

[0108] In addition, it should be noted that the four-point coordinates of the rectangle mentioned in this embodiment refer to the coordinate axes coordinates in the X and Y planes.

[0109] In addition, this embodiment also provides a readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the character encoding recognition method described above.

[0110] Since the technical improvement and the achieved technical effects of this readable storage medium are the same as those of the character encoding recognition method, the explanation of this readable storage medium will not be repeated here.

[0111] In addition, this embodiment also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the character encoding recognition method described above.

[0112] Since the technical improvements and achieved technical effects of this computer device are the same as those of the character encoding recognition method described above, this computer device will not be further explained.

[0113] In this way, in this embodiment, by combining an advanced deep learning model algorithm and a comparative vision algorithm, high accuracy and high robustness of character encoding detection are achieved.

[0114] Combined with data augmentation techniques and various post-processing algorithms, such as affine correction, secondary recognition, orientation classification, character merging, and standard character libraries, high-accuracy recognition of character encodings for workpieces in the harsh environment of shipyards is achieved. More importantly, for the problem of insufficient industrial scenario images, in this embodiment, through the above method, both images and labels are enhanced simultaneously, solving the problem of insufficient number of training images and greatly shortening the annotation cycle. In this way, compared with the detection and recognition of character encodings in existing vision algorithms and the detection and recognition of encodings relying solely on deep learning models, the detection accuracy of the character encoding recognition method in this embodiment is significantly improved.

[0115] In addition, the specific implementation of deploying a deep learning model on an industrial control computer in a complex environment such as a shipyard is also realized. Other solutions based on traditional vision algorithms or pure deep learning cannot meet such high accuracy requirements.

[0116] Although the present disclosure is disclosed as above, the protection scope of the present disclosure is not limited thereto. Without departing from the spirit and scope of the present disclosure, those skilled in the art can make various changes and modifications, and these changes and modifications will all fall within the protection scope of the present invention.

Claims

1. A character encoding recognition method, characterized in that: include: Through the transfer learning strategy, the detection model and the recognition model are trained separately using the training character encoding data; Converting the trained detection model and recognition model into corresponding detection reasoning model and recognition reasoning model respectively; Collecting a character-coded image at the workpiece; Using the detection inference model, obtaining encoding position plane posture information of multiple collected character strings in the character encoding image; Using the recognition inference model, based on the encoding position plane posture information of the plurality of collected character strings, obtaining information recognition results of the plurality of collected character strings; The information recognition results of the plurality of collected character strings are compared with all standard character strings in the standard character library to obtain the optimal comparison result for each of the collected character strings.

2. The character encoding recognition method according to claim 1, characterized in that: Through the transfer learning strategy, the detection model and the recognition model are trained separately using the training character encoding data, which also includes: Collect original character encoding data; Marking the original character encoding data; Performing enhancement and amplification processing on the marked original character encoding data to obtain the training character encoding data; The enhancement and amplification processing includes at least one of affine transformation, scaling, brightness increase and decrease, and expansion and corrosion.

3. The character encoding recognition method according to claim 1, characterized in that: Through the transfer learning strategy, the detection model and recognition model are trained separately using the training character encoding data, including: Adjust parameters, adopt transfer learning strategy, and combine the training character encoding data to train the DBNet character detection model in PaddleOCR; Adjust parameters, adopt transfer learning strategy, and combine the training character encoding data to train the SVTR-PPLCNetV3 character recognition model in PaddleOCR; Among them, the detection model is the DBNet character detection model, and the recognition model is the SVTR-PPLCNetV3 character recognition model.

4. The character encoding recognition method according to claim 1, characterized in that: Converting the trained detection model and recognition model into corresponding detection reasoning model and recognition reasoning model respectively includes: The detection model and the recognition model are deployed in the X64-CPU using the Paddle deep learning inference library.

5. The character encoding recognition method according to any one of claims 1 to 4, characterized in that: Using the detection inference model, obtaining the encoding position plane posture information of multiple collected character strings in the character encoding image includes: Using the detection reasoning model to detect the four-point coordinates of the rectangular box where the plurality of collected character strings are located; Processing the four-point coordinates of the rectangular boxes where the multiple collected character strings are located to obtain the corresponding areas of multiple detected character regions; Setting external parameters to filter false detections on the areas of the plurality of detected character regions; According to the angle between the longest character string among the plurality of collected character strings and the horizontal direction, affine transforming the longest character string among the plurality of collected character strings into a horizontal image; Performing secondary detection on the horizontal image to obtain coordinate information of the horizontal image; The coordinate information of the horizontal image is subjected to an inverse transformation check.

6. The character encoding recognition method according to claim 5, characterized in that: Using the recognition inference model, based on the encoding position plane posture information of the plurality of collected character strings, obtaining the information recognition results of the plurality of collected character strings comprises: According to the four-point coordinates of the rectangular frame where the multiple collected character strings are located and the corresponding areas of the multiple detected character regions, the corresponding area images where the multiple collected character strings are located are segmented from the character code image; Using the recognition inference model, respectively score the original images of the corresponding area images where the multiple collected character strings are located and the rotated images thereof after being rotated 180 degrees; Performing histogram equalization enhancement on the original image and / or the rotated image with a score lower than a set score, and scoring them again; The high-scoring ones of the original images and the rotated images, as well as the ones with scores higher than the set scores, are selected as the information recognition results of the collected character strings.

7. The character encoding recognition method according to claim 6, characterized in that: Comparing the information recognition results of the plurality of collected character strings with all standard character strings in the standard character library to obtain the optimal comparison result of each collected character string includes: Design a string merging algorithm based on the character distribution characteristics of shipyard artifacts; Using the string merging algorithm and setting an external parameter distance threshold, a plurality of the collected strings are merged; According to the information recognition results of the multiple collected character strings and the angle between the longest character string among the multiple collected character strings and the horizontal direction, the encoding position plane posture information of the merged multiple collected character strings is calculated.

8. The character encoding recognition method according to any one of claims 1 to 4, characterized in that: Comparing the information recognition results of the plurality of collected character strings with all standard character strings in the standard character library to obtain the optimal comparison result of each collected character string includes: Based on the optimal comparison result of each collected character string, the character string encoding information of each collected character string is output.

9. A readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the character encoding recognition method as described in any one of claims 1 to 8 is implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the character encoding recognition method according to any one of claims 1 to 8 is implemented.

Images