A container number intelligent identification method for guiding error correction by a detection frame edge expansion strategy and ISO 6346 rules

Abstract

The application discloses a kind of container number intelligent identification methods for detecting frame extension strategy and ISO 6346 rule guided error correction, belong to computer vision, intelligent port and optical character recognition field;The method comprises: the container number region in container image is positioned using target detection model;According to the size of detection frame, the container number region is proportionally extended, and the recognition region containing complete character context information is constructed;The image after extension is input into visual language model to identify container number, and obtain original recognition result;In combination with ISO standard, the recognition result is subjected to rule inference and error correction operation, and the final container number recognition result is output;The present application breaks through the character truncation problem caused by the boundary error of target detection frame, improves the integrity of container number region information;Break through the application mode that ISO 6346 rule is only used for result verification, realize the field rule knowledge participates in the correction and inference process of recognition result, solve the problem of visual language model in industrial scene recognition.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, smart ports and optical character recognition technology, and specifically relates to a method for intelligent identification of container numbers that integrates target detection, detection box expansion, visual language model recognition and ISO 6346 rule-guided error correction. Background Technology

[0002] Container numbers are unique identifiers for containers in international logistics, port handling, yard management, customs supervision, and multimodal transport. According to the ISO 6346 international standard, standard container numbers typically consist of four letters and seven numbers, with the last digit being a check digit calculated from the first ten characters. The accuracy of container number identification directly impacts the reliability of port automation operations, intelligent gate management, yard scheduling, and logistics information management. Existing container number identification methods typically employ a two-stage process: object detection + optical character recognition (OCR). Specifically, an object detection model first locates the container number region in an image, and then the image, cropped from the detection bounding box, is input into an OCR model for character recognition. While this method is effective in typical horizontal text scenarios, it suffers from the following shortcomings in real-world port, terminal, and yard environments. First, object detection models typically optimize the overlap between the detection bounding box and the labeled bounding box, focusing on boundary tightness in a detection sense. However, in container number identification tasks, the optimal boundary in a detection sense is not necessarily the optimal input region in a recognition sense. If the detection box is too tight, the first and last characters of the container number, character edge strokes, or check digit area may be cropped, causing subsequent recognition failures. Secondly, container numbers in real-world scenarios can be arranged horizontally, vertically, or in two rows. Traditional OCR methods are typically more suitable for horizontal single-line text. For vertically arranged, multi-line, or multi-column container numbers with complex reading sequences, problems such as character reading order errors, missing characters, or failed character recombination can easily occur. Thirdly, while visual language models possess strong image understanding and text generation capabilities and can handle character recognition problems in complex spatial layouts to some extent, their output may contain irrelevant text, repeated characters, missing characters, confusion between letters and numbers, abnormal lengths, and check digit errors. For industrial recognition tasks like container numbers, which have strict encoding rules, relying solely on the visual model output cannot guarantee that the results conform to the standard encoding format. Furthermore, in existing methods, the ISO 6346 rule is usually only used for final result verification, i.e., judging whether the recognition result conforms to the standard, without fully utilizing the prior knowledge in ISO 6346 regarding character position, encoding format, and check digit calculation to actively correct erroneous recognition results. Therefore, existing technologies still suffer from problems such as low recognition accuracy, insufficient anti-interference ability, and difficulty in recovering erroneous results in complex industrial scenarios. Summary of the Invention

[0003] The purpose of this invention is to address the problems in existing container number recognition technologies, such as detection box boundary errors causing character truncation, traditional OCR's inability to adapt to complex layouts, insufficient stability of visual language model output, and the failure of ISO 6346 rules to actively participate in error correction. This invention provides a container number intelligent recognition method that combines a detection box expansion strategy with ISO 6346 rules to guide error correction.

[0004] To achieve the above objectives, the present invention provides the following technical solution.

[0005] This invention provides a method for intelligent identification of container numbers guided by a detection frame expansion strategy and ISO 6346 rules, comprising the following steps: First, acquire a container image containing the container number and input the container image into the container number recognition system.

[0006] Second, the container number region in the container image is detected using an object detection model to obtain the location information of the detection box corresponding to the container number region. When the same container number is detected as multiple local detection boxes, the multiple detection boxes are merged to form the smallest bounding box that contains the complete container number region.

[0007] Third, a recognition-guided ROI expansion mechanism is constructed based on the width and height of the detection box. The detection box is expanded according to preset horizontal and vertical expansion ratios to generate an expanded ROI region containing the complete character structure and necessary contextual information. This expansion mechanism compensates for deviations in the target detection box boundary, reducing the risk of character truncation due to an overly tight detection box.

[0008] Fourth, an extended ROI image is obtained by cropping from the original image based on the extended boundary coordinates, and the extended ROI image is input into the visual language model; by using prompt words containing ISO 6346 format constraints and character reading order constraints, the visual language model is guided to identify the container numbers in the extended ROI image and output the original container number recognition result.

[0009] Fifth, a rule-guided error correction mechanism is constructed based on the ISO 6346 international standard coding rules. The original container number recognition results output by the visual language model are normalized, duplicate character illusion repaired, position-aware character mapping and check bit reconstruction are performed, and finally the container number recognition results that conform to the ISO 6346 coding rules are output.

[0010] Compared with the prior art, the present invention has the following beneficial effects: I. This invention proposes a recognition-oriented ROI expansion mechanism, which does not directly use the compact detection box output by the target detection model as the recognition input, but appropriately expands the edge of the detection box, so that characters that may be clipped near the edge of the box number are re-included in the recognition area, thereby improving the integrity of the box number character structure.

[0011] Second, this invention uses a visual language model to recognize extended ROI images, and utilizes its visual understanding and language reasoning capabilities to handle complex box number layouts such as horizontal, vertical and two-row layouts, thereby improving the adaptability to unconventional reading order scenarios.

[0012] Third, this invention introduces box number format constraints and character reading order prompts in the visual language model recognition stage, enabling the model to combine ISO 6346 format requirements and spatial structure information such as vertical and multi-column layouts when generating results, thereby reducing character reading order errors.

[0013] Fourth, this invention transforms the ISO 6346 rule from a traditional result verification tool into an active error correction mechanism. Through output normalization, duplicate character repair, position-aware character mapping, and check bit reconstruction, it corrects format errors, duplicate characters, alphanumeric confusion, and check bit errors in the output of the visual language model.

[0014] Fifth, this invention forms a complete technical process of "target detection - detection box expansion - visual language model recognition - ISO 6346 rule-guided error correction", which can improve the accuracy, robustness and engineering applicability of container number recognition in complex port scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope.

[0016] Figure 1 This is a schematic diagram of the overall process of the intelligent identification method for container numbers of the present invention; Figure 2 This is a schematic diagram illustrating the target detection frame merging and box number area positioning examples of the present invention; Figure 3 This is a schematic diagram of the error correction process guided by the ISO 6346 rule in this invention; Figure 4 This is a schematic diagram illustrating an example of the ISO 6346 rule-guided error correction process in this invention. Figure 5 This is a schematic diagram showing the training and validation results of the YOLOv11n detection model. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the described embodiments are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the core idea of ​​the present invention, those skilled in the art can make appropriate adjustments to the specific model type, parameter settings, or implementation method. Example

[0018] This embodiment provides a method for intelligent container number recognition based on a detection bounding box expansion strategy and ISO 6346 rule-guided error correction. The method mainly includes five parts: image input, container number region detection, ROI expansion, visual language model recognition, and ISO 6346 rule-guided error correction.

[0019] Step 1: Obtain images of the container.

[0020] Acquire container images from port, dock, yard, or logistics transportation scenarios. The container images contain container numbers conforming to ISO 6346 coding rules. Because container images in real-world scenarios may be affected by factors such as shooting angle, lighting changes, target scale, background interference, dirt, occlusion, or reflections, directly identifying container numbers from the entire image is easily affected by complex background interference.

[0021] In one specific embodiment, the data used comes from a publicly available online dataset of container code images, totaling 2981 samples. Of these, 2000 images are used for model building, 600 for parameter validation, and 381 for final evaluation. The test samples cover three common container number layouts: horizontal, vertical, and two-row, which can be used to verify the adaptability of the method of this invention to the recognition of container numbers with different spatial arrangements.

[0022] Step 2: Detection of container number area.

[0023] The input container image is fed into the object detection model, which outputs the coordinates of the detection bounding boxes for the container number regions in the image. Let the container number detection bounding boxes output by the object detection model be represented as: .

[0024] in, , This indicates the coordinates of the top-left corner of the detection box. , This indicates the coordinates of the bottom right corner of the detection box.

[0025] In real-world images, the same container number may be detected as multiple local regions due to vertical arrangement, two-row arrangement, or partial occlusion. In such cases, if the object detection model outputs multiple bounding boxes belonging to the same container number, these boxes are merged into a single minimum bounding box. Let these multiple bounding boxes be:

[0026] The coordinates of the merged minimum bounding box are: ; ; ;

[0027] This merging process integrates the scattered detected box number character regions into a unified recognition region, providing complete input for subsequent ROI expansion and visual language model recognition.

[0028] In one specific embodiment, the target detection model uses YOLOv11n. This model can balance detection accuracy and lightweight deployment requirements in the container number region detection task. The detection results are used to obtain the spatial location of the container number region, but the ultimate goal of this invention is not simply to pursue the optimal IoU of the detection bounding box, but to obtain an input region that is more suitable for container number recognition.

[0029] Step 3: Identify and expand the ROI.

[0030] Object detection models typically output relatively compact bounding boxes, aiming to accurately locate container number regions. However, for container number recognition, overly tight bounding boxes may clip character edges, first and last characters, or check digits. Therefore, this invention constructs a recognition-oriented ROI expansion mechanism based on the bounding box.

[0031] Set detection frame The width and height are as follows: ; .

[0032] Let the horizontal expansion ratio be... The vertical expansion ratio is The original image width is The height is Then the coordinates of the expanded detection box are: ; ; ;

[0033] The expanded detection box is represented as follows:

[0034] The expanded ROI image is obtained by re-cropping from the original image based on the coordinates B' of the expanded detection box.

[0035] The purpose of this ROI expansion mechanism is to compensate for the uncertainty of the detection box boundaries, allowing characters that may be truncated near the edge of the detection box to re-enter the recognition area. If the expansion ratio is too small, it may not adequately solve the character truncation problem; if the expansion ratio is too large, it may introduce irrelevant background text or other box-shaped markings, interfering with the visual language model's recognition. Therefore, the expansion ratio should be set according to the recognition performance and the degree of background redundancy.

[0036] In one specific embodiment, a symmetrical edge expansion strategy is adopted, namely... Through the expansion ratio ablation experiment, the following was selected: This is the default border expansion configuration. This setting improves the recognition integrity of boundary characters and complex layout characters while keeping the ROI area compact.

[0037] Step 4: Visual language model recognition.

[0038] The extended ROI image is input into a visual language model for container number recognition. The visual language model can directly receive image input and generate text output without explicit character segmentation. Compared to traditional OCR models, the visual language model can combine the appearance, spatial arrangement, and language prompts of characters in the image to perform more flexible reading order inference for vertical, multi-column, or two-row container numbers.

[0039] Let the visual language model be denoted as The expanded ROI image is denoted as ROI, and the recognition prompt word is denoted as... Then the original box number recognition result output by the visual language model is:

[0040] in, This represents the original recognition string generated by the visual language model.

[0041] To guide the visual language model in reading characters according to the standard format and actual layout of container numbers, this invention constructs a recognition prompt that includes format constraints and reading order constraints. The recognition prompt includes: Identify container numbers in an image; Following the ISO 6346 standard, output an 11-digit number consisting of four letters and seven numbers; If the box number is in vertical format, read it from top to bottom; If the box number is in a multi-column vertical format, read the left letter column first, then read the right number column; Output only the 11-digit container number code.

[0042] In this invention, layout category is not used as an additional model input condition, nor is separate layout classification required before recognition. The visual language model generates recognition results based on the spatial arrangement of characters in the extended ROI image and the reading order constraints in the prompt words. Layout classification can be used for subsequent experimental statistics and performance analysis, but it does not participate in the recognition process itself.

[0043] In one specific embodiment, the visual language model employs Qwen2-VL-7B-Instruct. This model possesses image understanding and text generation capabilities, and can handle ROI images with different aspect ratios and layouts. For dynamic resolution visual language models, a moderately enlarged ROI region can retain more character boundaries and local contextual information, which is beneficial for improving the completeness of box number recognition.

[0044] Step 5: ISO 6346 rules guide error correction.

[0045] The raw recognition results output by the visual language model may contain spaces, punctuation marks, line breaks, explanatory text, character repetitions, missing characters, alphanumeric confusion, or checksum errors. To improve the structural reliability of the output results, this invention constructs a rule-guided error correction mechanism based on the ISO 6346 encoding rules.

[0046] The ISO 6346 standard container number consists of four letters and seven numbers, with the eleventh digit being a check digit. This invention utilizes this structural rule to actively reason and correct errors in the output of the visual language model, rather than simply judging the validity of the results.

[0047] The ISO 6346 rule-guided error correction includes the following steps.

[0048] Step 5.1, output normalization.

[0049] Normalization is performed on the raw strings output by the visual language model. Specifically, this includes: Remove non-alphanumeric characters from the original string; Convert all letters to uppercase; Use the processed string as the candidate string; When the length of a candidate string exceeds 11 characters, the first 11 characters are truncated as the candidate structure.

[0050] Output normalization can eliminate spaces, punctuation marks, line breaks, and irrelevant explanatory text in the output of visual language models, providing a unified input for subsequent rule-based reasoning.

[0051] Step 5.2, Duplicate Character Illusion Repair.

[0052] Visual language models may produce consecutive repeated characters during the generation process. Since two consecutive identical characters in a container number may be valid, while three or more consecutive identical characters are more likely to originate from repeated output errors in the generative model, this invention focuses on handling three or more consecutive identical characters.

[0053] Specifically, the process begins by scanning the first character of the candidate string. If three or more consecutive identical characters are found, one of the duplicate characters is deleted to form a revised candidate string. The revised candidate string is then checked for consistency according to the ISO 6346 checksum rules. If the candidate string after deleting duplicate characters satisfies the ISO 6346 checksum consistency, the revised candidate string is adopted; otherwise, the original candidate string is retained.

[0054] This process does not arbitrarily delete duplicate characters, but uses ISO 6346 verification as an acceptance condition, thereby reducing the risk of erroneous correction.

[0055] Step 5.3, position-aware character mapping.

[0056] According to ISO 6346, the first four digits of a container number should be letters, and the fifth to eleventh digits should be numbers. Visual language models may confuse similar-looking letters and numbers, for example, mistaking the letter O for the number 0, or the number 5 for the letter S. To address this problem, this invention performs position-aware character mapping based on character position constraints.

[0057] For the first four digits, if the recognition result is a number but the number is visually similar to a certain letter, it is mapped to the corresponding letter. For example, the numbers 0, 1, 5, 8, and 2 can be mapped to the letters O, I, S, B, and Z, respectively.

[0058] For the 5th to 11th positions, if the recognition result is a letter but the letter is visually similar to a certain number, it is mapped to the corresponding number. For example, the letters O, I, Z, S, and B can be mapped to the numbers 0, 1, 2, 5, and 8, respectively.

[0059] This location-aware character mapping allows for the use of ISO 6346 coding structure constraints to correct recognition errors caused by visually similar characters.

[0060] Step 5.4, reconstruct the check bit.

[0061] After completing output normalization, duplicate character illusion repair, and position-aware character mapping, the eleventh bit of the candidate string is further checked and reconstructed according to the ISO6346 check bit calculation rules.

[0062] Let the first ten characters of the candidate box number be:

[0063] According to ISO 6346, each character is mapped to its corresponding numerical value v. i And perform a weighted sum based on character position:

[0064] according to Calculate the check digit .

[0065] when At that time, the check bit will be... Set to 0; otherwise, check bit for The calculated check digit is used as the eleventh character to generate the final container number identification result.

[0066] By reconstructing the check bits, erroneous or missing check bits in the visual language model output can be corrected, so that the final result meets the ISO 6346 encoding rules. Example

[0067] To verify the effectiveness of the method of the present invention, in one specific embodiment, an experiment was conducted using a publicly available container coding image dataset. This dataset contains a total of 2981 samples, of which 2000 were used for model building, 600 for parameter validation, and 381 for final evaluation. The final evaluation samples included three container number layouts: horizontal, vertical, and two-row.

[0068] In the target detection stage, YOLOv11n was used to locate the container number regions. In the test set, the YOLOv11n detection module achieved a precision of 99.4%, a recall of 99.8%, a mAP@0.5 of 99.3%, and a mAP@0.5:0.95 of 80.7%. These results demonstrate that the target detection model can locate the vast majority of container number regions, providing a reliable foundation for subsequent ROI expansion and identification.

[0069] During the recognition phase, comparative analysis shows that ROI expansion significantly improves the visual language model. With ISO6346 rule correction removed and only basic output normalization retained, Qwen2-VL achieved a complete box number recognition accuracy of 34.91% without ROI expansion, which increased to 69.82% after adding ROI expansion. This demonstrates that recognition-oriented ROI expansion effectively compensates for bounding box boundary errors, improving character integrity and recognition stability.

[0070] Furthermore, after employing a complete process including object detection, ROI expansion, visual language model recognition, and error correction guided by ISO 6346 rules, the overall accuracy of complete box number recognition reached 98.16%. This result demonstrates that the method of the present invention can adapt to different box number layouts.

[0071] Furthermore, ISO 6346 rule-guided error correction can significantly improve the structural reliability of the final output. Through output normalization, duplicate character illusion repair, position-aware character mapping, and check digit reconstruction, this invention can transform the raw output of the visual language model into container number results conforming to the ISO 6346 standard, reducing format errors, character obfuscation, and check digit errors in the output of generative models.

[0072] As can be seen from the above embodiments, the present invention improves the recognition input area through the detection box expansion strategy, processes complex spatial layouts through a visual language model, and improves the output standardization and accuracy through error correction guided by ISO 6346 rules. It can effectively improve the accuracy and robustness of container number recognition in complex industrial scenarios.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any equivalent substitutions, improvements, or modifications made to the target detection model, visual language model, edge expansion ratio, cue word expression method, or ISO 6346 rule error correction implementation method within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of container numbers guided by a detection frame expansion strategy and ISO 6346 rules, characterized in that, The intelligent identification method for container numbers includes the following steps: (1) Collect images of containers in ports, docks, storage yards and logistics transportation scenarios, and input the images into the container number recognition system; (2) Use the target detection model to detect the box number region in the input image and obtain the position coordinate information of the detection box corresponding to the box number; (3) Construct an ROI expansion mechanism based on the width and height of the detection box, expand the boundary of the detection box according to the preset horizontal expansion ratio and vertical expansion ratio, and generate an expanded recognition area containing complete character information and context information; (4) Based on the boundary coordinates after expansion, the original image is re-cropped to obtain the expanded ROI image, and the ROI image is input into the visual language model; the visual language model is input with recognition prompts containing container number format constraints and character reading order constraints, so that the visual language model can recognize the container number characters in the expanded ROI image and output the original container number recognition result. (5) The original container number identification result is subjected to rule-guided error correction processing based on the ISO 6346 international standard coding rules. The rule-guided error correction processing includes output normalization, duplicate character illusion repair, position-aware character mapping and check bit reconstruction, and outputs the final container number identification result that conforms to the ISO 6346 coding rules.

2. The intelligent container number identification method according to claim 1, characterized in that, In step (2), when the target detection model outputs one or more container number area detection boxes in the same container image, if there are multiple container number area detection boxes, the multiple container number area detection boxes are merged to obtain the smallest outer detection area containing the complete container number area, and the smallest outer detection area is used as the initial container number area for subsequent ROI expansion. Let the plurality of detection boxes be represented as: in, , Indicates the first The coordinates of the top left corner of the detection box , Indicates the first The coordinates of the bottom right corner of each detection box are used to represent the smallest bounding detection area after merging: ； ； ； in, This is the merged box number detection frame.

3. The intelligent container number identification method according to claim 1, characterized in that, In step (3), the box number detection frame is set as follows: The detection box width W and height H are respectively: ； Let the horizontal expansion ratio be... The vertical expansion ratio is The original container image width is The height is Then the coordinates of the expanded detection box are: ; ; ; The expanded detection box is as follows: . And based on the coordinates of the expanded detection frame An expanded ROI image is obtained by cropping from the original container image; The horizontal expansion ratio and longitudinal expansion ratio It adopts a symmetrical expansion setting, and Preferably, .

4. The intelligent container number identification method according to claim 1, characterized in that, In step (4), the expanded ROI is input into the visual language model for recognition. Let the visual language model be denoted as The extended ROI image is denoted as The recognition prompt words are recorded as Then the original box number recognition result output by the visual language model is: in: ; Indicates the first in the recognition result One character; Indicates the length of the recognition result. Further construct the recognition prompts: "Recognize the container number in the image"; "Follow the ISO 6346 standard (4 letters + 7 numbers)"; "If it is vertical, please read from top to bottom"; "If it is a multi-column vertical arrangement, please read the left letter column first, and then read the right number column"; "Output only the 11-digit number". To reduce the probability of the visual language model generating irrelevant text, container number recognition prompts are constructed to guide the model to output only the container number code. The visual language model output may exhibit issues such as missing characters, repeated characters, confusion between letters and numbers, abnormal length, and formatting errors; therefore, further reasoning and correction based on ISO 6346 rules are necessary. The visual language model's recognition process does not use the individually predicted layout category as input. Instead, it uses the reading order constraint in the recognition prompt to enable the visual language model to generate the original box number recognition result based on the spatial arrangement of characters in the extended ROI image.

5. The intelligent container number identification method according to claim 1, characterized in that, The output normalization in step (5) includes: Remove non-alphanumeric characters from the original box number identification results; Convert all letters to uppercase. The normalized strings are then organized into candidate strings for subsequent rule reasoning; When the length of the normalized string exceeds 11 bits, 11-bit candidate structures are extracted from the normalized string according to the ISO 6346 encoding length requirements.

6. The intelligent container number identification method according to claim 1, characterized in that, The repetitive character illusion repair in step (5) includes: Perform a continuous character scan on the normalized candidate string; When three or more consecutive identical characters are found, one of the duplicate characters is deleted to form a correction candidate string; The modified candidate string is validated according to the ISO 6346 check bit rules; If the corrected candidate string after deleting duplicate characters satisfies the ISO 6346 checksum consistency, the corrected candidate string is used as the result of subsequent processing; otherwise, the candidate string before deletion is retained.

7. The intelligent container number identification method according to claim 1, characterized in that, The position-aware character mapping in step (5) includes: According to the ISO 6346 coding rules, the first to fourth characters of the container number are constrained to be alphanumeric characters, and the fifth to eleventh characters are constrained to be numeric characters. For characters in positions 1 through 4 that are identified as numbers but visually similar to letters, they are mapped to their corresponding letters according to visual similarity. For characters in positions 5 through 11 that are identified as letters but visually similar to numbers, they are mapped to their corresponding numbers according to visual similarity. The numbers 0, 1, 5, 8, and 2 can be mapped to the letters O, I, S, B, and Z, respectively, and the letters O, I, Z, S, and B can be mapped to the numbers 0, 1, 2, 5, and 8, respectively.

8. The intelligent container number identification method according to claim 1, characterized in that, The parity bit reconstruction in step (5) includes: Extract the first ten characters of the candidate string after output normalization, duplicate character illusion repair, and position-aware character mapping; The first ten characters are mapped to their corresponding numerical values ​​according to the ISO 6346 encoding rules. ; The numerical value is determined according to the character position. We perform a weighted summation and get: ; according to Calculate the check digit ; when At that time, the check bit will be... Set to 0; otherwise, check bit for ,Right now: ; The calculated check digit is used as the 11th character to generate the final container number identification result.

9. The intelligent container number identification method according to claim 1, characterized in that, The ROI expansion mechanism is used to compensate for the boundary deviation of the target detection box, so that the character structure that is truncated or may be truncated near the edge of the detection box is re-included in the expanded ROI image, thereby improving the visual language model's recognition completeness of the first and last characters of the container number, vertical container number and two-line container number.

10. The intelligent container number identification method according to claim 1, characterized in that, The visual language model is a multimodal visual language model that can receive image input and generate text output. The visual language model generates the original box number recognition result based on the character appearance information, spatial arrangement information and reading order constraints in the recognition prompt in the extended ROI image.

Images