Tire mold character automatic detection method based on image processing

Abstract

The application discloses a tire mold character automatic detection method based on image processing and concretely relates to the technical field of image processing and mold detection; under fixed illumination conditions, a mold image is collected by an industrial camera; the image is preprocessed and regionally segmented, and a candidate character region is generated based on a deep convolution feature extraction network; edge density analysis and minimum circumscribed rectangle fitting are performed on the candidate region, the character arrangement direction is judged, and image direction normalization is completed; the processed character image is input into a pre-trained character recognition model for recognition and compared with a mold character template in a production planning; if the recognition result is inconsistent with the template or the confidence degree is lower than a set threshold, the abnormal region is marked and an alarm information is generated; the application has the advantages of high recognition precision, accurate direction judgment, fast feedback response and the like, can realize automatic recognition and intelligent verification of mold characters and is suitable for mold installation quality control in a tire manufacturing site.

Description

Technical Field

[0001] This invention relates to the field of image processing and mold inspection technology, specifically to an automatic character detection method for tire molds based on image processing. Background Technology

[0002] Currently, in the tire production process, to ensure the traceability and consistency of finished tire information, tire mold surfaces need to be engraved with identifying characters including specifications, production dates, and manufacturer codes. These characters are mostly engraved in recessed form and are distributed across multiple irregular curved areas on the inner wall of the mold. This presents challenges such as surface tilt, glare interference, and dust accumulation, leading to the following shortcomings in practical applications of traditional OCR (Optical Character Recognition) based character detection methods: Image acquisition accuracy is highly dependent: traditional methods are difficult to achieve fast and clear character image acquisition at the mold installation site, especially when the mold has a large curvature or the characters are shallowly engraved, the recognition accuracy drops significantly. Character detection directionality is uncertain: Existing methods have difficulty determining the forward arrangement relationship of characters, requiring manual verification of the direction one by one, which increases labor costs; Character positioning requires a lot of manual intervention: most current systems cannot automatically determine the specific mold area where the character is located, and inspectors need to compare the template images one by one; Insufficient robustness in character recognition: When there are interfering factors such as scratches, oil stains, and dust accumulation on the mold surface, the accuracy of character recognition drops significantly, seriously affecting the detection efficiency and mold installation quality.

[0003] Given the above issues, especially during mold installation, it is necessary to achieve rapid character detection and positioning and feedback on whether it matches the current production schedule. If manual judgment is still relied upon, it will not only be inefficient, but also pose a risk of mold assembly errors, product scrap, or even batch quality accidents due to misjudgment. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic character detection method for tire molds based on image processing, so as to overcome the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic character detection method for tire molds based on image processing, comprising: The image of the tire mold is acquired by an industrial camera located at the mold inspection station under fixed lighting conditions. The mold includes at least one area for engraving characters on a curved surface. The mold image is segmented into regions, and a candidate region is generated for regions in the image that may contain characters based on a deep convolutional feature extraction network, resulting in several candidate character regions. Geometric morphology correction is performed on each candidate character region. The character arrangement direction is determined based on the character edge density and curvature characteristics. Direction normalization is then performed to obtain the character forward image. Input the forward image of the character into the pre-trained character recognition model to identify the content of each character and compare it with the corresponding mold character template data in the current production schedule to generate character matching results; If there are character items whose character recognition results do not match the template or whose confidence level is lower than the set threshold, the character area is marked as an abnormal area and a corresponding alarm message is generated. The alarm information, along with the original image and character position coordinates, is stored in the mold detection database. The abnormal detection results are then fed back to the mold installation terminal through the MES system to provide on-site abnormal alerts.

[0006] Preferably, the step of generating candidate character regions includes: The mold image is subjected to grayscale normalization and Gaussian filtering to enhance the contrast between character edges and background; Based on the image pyramid structure, a sliding window scan is performed on multi-scale images to initially screen out image sub-regions with character texture features; Feature vectors are extracted from image sub-regions using a feature extraction model built on a deep convolutional neural network, and then input into a classifier for character region discrimination. The coordinates of the image sub-regions identified as character regions are used as candidate character regions.

[0007] Preferably, the step of performing geometric morphological correction on each candidate character region includes: Edge extraction processing is performed on the candidate character region. The Canny algorithm is used to extract the character outline and calculate the edge density distribution of the outline pixels in the horizontal and vertical directions. The main direction of the character is determined based on the direction of maximum edge density, and the angle of the main axis of the character arrangement is extracted by combining the minimum bounding rectangle fitted by the character outline within the region. Affine transformation correction is performed on the character region image based on the principal axis direction angle to align the character arrangement direction with the horizontal direction of the image. The corrected image is then processed for edge completion and size normalization to generate a character-oriented image region.

[0008] Preferably, the step of calculating the edge density distribution of contour pixels in the horizontal and vertical directions includes: The extracted contour image is scanned by row and column respectively, and the number of edge pixels in each row and column is counted. Divide the number of edge pixels in each row by the total number of pixels in that row to obtain the horizontal edge density sequence; divide the number of edge pixels in each column by the total number of pixels in that column to obtain the vertical edge density sequence. The mean and variance of the edge density sequences in the horizontal and vertical directions are calculated respectively to measure the central tendency and directional consistency of the edge distribution. By comparing the average density and variance of the two directions, the main direction of the edge distribution is determined, which helps to determine the main axis angle of the character arrangement direction.

[0009] Preferably, the step of extracting the angle of the main axis of character arrangement includes: By comparing the average values ​​of the horizontal and vertical edge density sequences in the candidate character region, the main direction of the edge density distribution is determined. Based on the determined main direction, the coordinate set of the pixel points of the character region outline is fitted, and the least squares method is used to fit its minimum bounding rectangle. Extract the angle between the long side of the minimum bounding rectangle and the horizontal direction of the image, and use it as the angle of the main axis direction of character arrangement; If the spindle direction angle exceeds the preset angle threshold range, the direction normalization process is triggered.

[0010] Preferably, the step of generating character matching results includes: The forward-facing character images that have undergone orientation and size normalization are input into the pre-trained character recognition model with uniform pixel resolution and grayscale range. A character recognition model based on a convolutional neural network is used to extract features and classify character images, and output the corresponding character category and its recognition confidence. Obtain the mold character template data corresponding to the current production schedule, and compare the character categories with the target characters in the template data one by one; Based on the consistency between the character recognition result and the template character, and whether the recognition confidence level meets the preset threshold conditions, a character matching result is generated.

[0011] Preferably, the industrial camera is mounted on a fixed bracket with multi-degree-of-freedom adjustment capability, and is illuminated by a ring-shaped light-emitting diode light source within a closed or semi-closed light cover, so that the curved surface engraved character area forms a stable light and dark contrast feature in the image.

[0012] Preferably, the generation of the character matching result includes: comparing the character recognition result with the corresponding mold character template data in the production schedule according to the spatial order of the character region in the mold image.

[0013] Preferably, when the character recognition result is inconsistent with the template character, or the character recognition confidence level is lower than the set threshold, the corresponding character area is marked as an abnormal area, and the abnormal type and abnormal location coordinates are recorded.

[0014] Preferably, after determining the main direction of the character, the minimum bounding rectangle is fitted to the set of character outline pixels, and the angle between the long side of the minimum bounding rectangle and the horizontal direction of the image is extracted as the angle of the main axis of the character arrangement; when the angle of the main axis of the character arrangement exceeds a preset angle threshold, affine transformation correction is performed on the candidate character region to align the character arrangement direction with the horizontal direction of the image.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs an automatic character detection method for tire molds, encompassing image acquisition, character candidate region extraction, orientation normalization, depth recognition, template comparison, anomaly detection, and on-site feedback. This method effectively addresses the challenge of character recognition under complex imaging conditions such as tilting, reflection, and shallow engraving. It employs a combination of deep convolutional feature extraction and support vector machine classification to achieve high-precision character region detection. By combining edge density analysis and geometric fitting methods, it enables automatic character orientation judgment and correction, significantly improving recognition accuracy and system stability. This solution eliminates the need for manual intervention to complete character recognition and template verification, greatly reducing labor intensity and the risk of misjudgment.

[0016] 2. This invention possesses a complete anomaly recognition and feedback mechanism, enabling accurate marking and real-time alarms for abnormal characters. Combined with the manufacturing execution system, anomaly information is synchronized to the mold installation terminal, and the detection images, character coordinates, and recognition results are stored in a structured manner in the mold inspection database, supporting subsequent historical queries and quality traceability. Through the integration of these functions, the system not only improves the inspection efficiency and automation level of the mold installation process but also achieves digital closed-loop management of the mold recognition process, possessing excellent industrial deployability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the automatic character detection method for tire molds based on image processing according to the present invention.

[0019] Figure 2 This is a flowchart of the candidate character region geometric shape correction method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] For examples, please refer to Figure 1 , 2 As shown in this embodiment, the automatic character detection method for tire molds based on image processing includes: The image of the tire mold is acquired by an industrial camera located at the mold inspection station under fixed lighting conditions. The mold includes at least one area for engraving characters on a curved surface.

[0022] In this invention, an image acquisition module is installed at the tire mold installation and inspection station. It is configured to acquire images of character areas on the mold surface during mold assembly, enabling subsequent image processing and character recognition. This step is the initial stage of the entire inspection process, and its image quality and acquisition accuracy directly affect the accuracy of subsequent recognition and the stability of the system.

[0023] Specifically, the image acquisition module includes at least one industrial camera, a fixed mounting bracket, a matching light source system, and an image acquisition control unit.

[0024] Industrial camera selection and installation location: A color industrial camera with a resolution of at least 5 megapixels is preferred, offering high image detail reproduction and low noise performance. To accommodate the curved surface structures of different mold models, the industrial camera is mounted on a bracket with multi-degree-of-freedom adjustment capabilities. The bracket is positioned above or to the side of the inspection station, allowing for rotation, pitch, and fine-tuning of the camera's position according to the mold's installation angle. This ensures the camera lens axis is as perpendicular as possible to the curved surface area where the target character is located, thereby reducing image distortion.

[0025] Fixed lighting design: To eliminate interference from ambient light variations in image acquisition, the image acquisition area is equipped with a closed or semi-closed light cover structure, and is fitted with a ring-shaped LED light source or a linear oblique side lighting system. The brightness, angle, and color temperature of the light source have been experimentally optimized to ensure a stable high-contrast effect in the image of the etched character area. Ideally, the character area should exhibit a contrast characteristic of bright edges and a dark bottom surface, facilitating subsequent character outline extraction.

[0026] Trigger Control and Image Acquisition: The industrial camera is communicatively connected to the control module of the detection system. Image acquisition is triggered by external sensors (such as mold positioning completion signals) or PLC signals. After the acquisition command is issued, the camera completes image capture according to preset parameters (exposure time, gain, white balance, etc.) and transmits the image to the host computer or image processing module via a gigabit Ethernet interface.

[0027] Character Area Characteristics and Adaptability Design: This invention is applicable to tire molds containing at least one engraved character area. This character area is typically located on the inner curved wall of the mold, with a metal surface. The characters are mostly incised, and the font size is between 2mm and 6mm. Considering the imaging challenges of such areas, including large curvature variations, strong reflections, and a high possibility of occlusion, this invention prioritizes the following factors in the image acquisition module design: Multi-camera collaborative acquisition: For molds with a wide range of character positions, the system can be configured with two or more cameras to acquire data from different areas, ensuring no blind spots.

[0028] Image stabilization algorithm: The image acquisition and control software performs inter-frame comparison and motion compensation on the original image to ensure image stability and accurate focus.

[0029] Through the above methods, the present invention can stably and clearly acquire images of the mold surface containing characters without affecting the mold installation efficiency, providing a high-quality image foundation for subsequent character region recognition and text content extraction.

[0030] The mold image is segmented into regions, and a candidate region is generated for regions in the image that may contain characters based on a deep convolutional feature extraction network, resulting in several candidate character regions.

[0031] After acquiring the tire mold image, in order to achieve high-precision character recognition, it is necessary to accurately segment the effective region containing the character from the overall image. The specific implementation method is as follows: First, the input color mold image is converted to grayscale. Using a weighted average method, the red, green, and blue channels are multiplied by 0.299, 0.587, and 0.114 respectively and then summed to form a single-channel grayscale image, ensuring that the character edges have a clear representation in terms of brightness distribution.

[0032] Subsequently, to eliminate high-frequency noise interference caused by uneven lighting or metal reflections from the mold, a two-dimensional Gaussian filter was used to smooth the image. The expression for the Gaussian kernel function used is: ; where σ is set to 1.5 and the kernel size is 5×5. The filtering operation uses this kernel function as the weight to weight the neighboring pixel values ​​at each pixel, thereby achieving local smoothing and highlighting character edges.

[0033] After processing, a preprocessed image with uniform grayscale distribution and intact edge features is obtained, providing a clean background for subsequent area scanning.

[0034] To address the issue of inconsistent character sizes in images, a multi-layered pyramid image structure is constructed. Starting with the original image as the first layer, three sub-images are generated through continuous downsampling using bilinear interpolation, with scaling ratios of 0.75, 0.5, and 0.25 respectively, resulting in a total of four image pyramid layers.

[0035] On each image layer, a fixed-size sliding window of 32 pixels × 32 pixels is used for region scanning, with a horizontal and vertical step size of 8 pixels. For each window image patch, the following features are extracted: Histogram of Oriented Gradients (HOG): Divide the window into 8 blocks, count the gradient direction distribution in each block, and construct a 64-dimensional feature vector; Gray-level co-occurrence matrix features: extracting three texture features—energy, contrast, and homogeneity—to characterize the local texture patterns of characters; Edge density: The Sobel operator is used to calculate the percentage of pixels at the edge of the window. Edges with a density exceeding a set threshold (e.g., 15%) are considered candidates.

[0036] The above features are combined into a preliminary 70-dimensional feature vector, which is used as input for character region candidate selection.

[0037] The candidate image regions selected through the sliding window are uniformly adjusted to a 32-pixel × 32-pixel input size and then input into a trained deep convolutional neural network to extract advanced semantic features.

[0038] The deep network structure is as follows: Layer 1: Convolutional layer, kernel size 3×3, number of channels 16, stride 1, activation function is ReLU; Layer 2: Max pooling layer, window size 2×2, stride 2; Layer 3: Convolutional layer, kernel size 3×3, number of channels 32, stride 1; Layer 4: Max pooling layer, window size 2×2, stride 2; Layer 5: Fully connected layer, with an output feature dimension of 128.

[0039] The network training uses a dedicated dataset containing character images and background images of tire molds, with a total of no less than 50,000 images. The character samples cover various fonts, depths, and lighting conditions. The training process uses the cross-entropy loss function as the objective function, and the Adam optimizer is selected as the optimization algorithm. The initial learning rate is set to 0.001, the total training epochs are 100, and the batch size is 128.

[0040] After training, the parameters of the convolutional network are frozen, and the output 128-dimensional feature vector is input into the support vector machine classifier for binary classification.

[0041] The support vector machine uses a radial basis function kernel with a penalty coefficient set to 1.0, and the ratio of positive to negative samples in the training data is 1:1. The classifier returns a label of 0 or 1, representing "non-character region" and "character region" respectively.

[0042] For a window image that is identified as a character region by the classifier, its position coordinates in the original image are recorded, including the horizontal coordinate of the top left corner x1, the vertical coordinate of the top left corner y1, the window width w, and the height h, forming a set of candidate character region coordinates.

[0043] To avoid redundant regions caused by overlapping sliding windows, a non-maximum suppression algorithm is used to filter the candidate set. A threshold of 0.6 is set for the intersection-union ratio (IU) of overlapping regions. When the IU of two candidate regions exceeds this threshold, only the region with the higher classification probability is retained, and the remaining overlapping regions are deleted. The final set of non-repeating regions is used as candidate regions for character detection and input into the next stage of directional normalization.

[0044] Geometric morphology correction is performed on each candidate character region. The character arrangement direction is determined based on the character edge density and curvature characteristics. Direction normalization is then performed to obtain the character forward image.

[0045] After extracting the candidate character regions, to further improve the accuracy and stability of character recognition, it is necessary to analyze the arrangement direction of each candidate character region and perform image orientation normalization. This embodiment calculates the character arrangement direction by combining edge density analysis with minimum bounding rectangle fitting, and performs affine transformation based on the principal axis angle to achieve character orientation alignment. The processing steps are as follows: For each candidate character region image, the Canny edge detection algorithm is used to extract the character contour edges. A low threshold of 50 and a high threshold of 150 are set. The Sobel operator is used for gradient calculation, and a double-threshold hysteresis connection method is used for edge connectivity. The Canny algorithm outputs a binary image, where edge pixel values ​​are 255 and non-edge pixel values ​​are 0.

[0046] After edge extraction, the binary edge image is traversed both horizontally and vertically. The number of edge pixels (with a pixel value of 255) in each row is counted and recorded as the edge pixel count for that row. This value is then divided by the total number of pixels in the row (equal to the image width) to calculate the horizontal edge density value. This process is repeated column-wise to calculate the vertical edge density value. The density values ​​from all rows are used to construct a horizontal edge density sequence, and the density values ​​from all columns are used to construct a vertical edge density sequence.

[0047] The statistical characteristics of the edge density sequences in both directions are calculated, including the mean (reflecting the overall density of the edges in that direction) and the variance (reflecting the dispersion of the edge distribution). The statistical calculations are expressed as follows: The average value is the sum of all edge density values ​​divided by the number of pixel rows or columns; Variance is the sum of squares of the differences between all density values ​​and their mean, divided by the total number of samples.

[0048] This process enables quantitative analysis of the edge concentration of character structures in the horizontal and vertical directions, providing a basis for determining the character arrangement direction.

[0049] The main character arrangement direction is determined by comparing the average edge density sequences in the horizontal and vertical directions: if the average edge density in the horizontal direction is higher than that in the vertical direction (the difference exceeds the set threshold of 0.05), the main character arrangement direction is determined to be horizontal; if the edge density in the vertical direction is higher, it is determined to be vertical.

[0050] Given the main direction, we further extract the set of coordinate points of all edge pixels and perform minimum bounding rectangle fitting on this set. The fitting process uses the least squares method to fit all edge pixels into a set of rotated rectangular bounding boxes, and then obtains the direction vector of the longest side.

[0051] Let the angle θ be the angle between the long side vector of the smallest bounding rectangle and the horizontal direction of the image. This angle is the angle of the main axis of the character arrangement. The angle range is set from 0 to 90 degrees. When θ is close to 0 or 90 degrees, it means that the character arrangement is relatively horizontal or vertical. If θ is a tilt angle (for example, between 15 and 75 degrees), then rotation correction is required.

[0052] When the angle θ of the principal axis of the character arrangement exceeds the set angle threshold range (e.g., 5 degrees), an affine transformation correction needs to be performed on the candidate character region to align its arrangement direction with the horizontal direction of the image. The affine transformation is constructed using a two-dimensional rotation matrix, with the angle parameter being a negative value of θ, used to rotate the characters to a horizontal state.

[0053] The affine transformation matrix is ​​expressed as: The image center is used as the rotation center to perform the matrix transformation, obtaining an orientation-aligned character image region. Boundary extrapolation uses an edge copying mode for padding to avoid truncation of character content.

[0054] After orientation alignment, character image areas may suffer from incomplete character edges or inconsistent image sizes due to rotation or cropping. To address this issue, the image edges are supplemented using a mirror expansion method, with the maximum expansion size not exceeding 20% ​​of the original region's side length.

[0055] Subsequently, the character image size was uniformly normalized to 64 pixels × 64 pixels, maintaining the aspect ratio, and any excess areas were filled with a grayscale value of 128. After processing, the character areas have a uniform orientation and size, facilitating accurate input to the subsequent character recognition model.

[0056] By sequentially processing the above steps, accurate determination of character arrangement direction and image geometric correction are achieved, significantly improving the robustness of the recognition model to mold characters in different directions. It is especially suitable for complex scenarios where characters are tilted, circularly distributed, or have uneven spacing.

[0057] The forward image of the character is input into the pre-trained character recognition model to identify the content of each character and compare it with the corresponding mold character template data in the current production schedule to generate character matching results.

[0058] After completing the orientation and size normalization of the character images, in order to automatically compare and verify the mold characters with the production schedule, it is necessary to recognize the content of the character images and perform consistency judgment in conjunction with the plan data. This embodiment constructs a deep convolutional neural network character recognition model and combines it with character template data to output the character matching results. The overall processing flow is as follows: Before inputting the character images, which have undergone orientation and size normalization, into the recognition model, a unified formatting process is first performed. The size of the character images is uniformly adjusted to 64 pixels × 64 pixels, and the pixel grayscale range is normalized to between 0 and 1, i.e., each pixel value is divided by 255. The images are uniformly converted to floating-point format to form the model input tensor, with a dimension format of [1, 1, 64, 64], where the first dimension is the batch dimension and the second dimension is the number of channels, representing a single-channel grayscale image. This step ensures that the input data format of the character recognition model is consistent, improving the model's recognition accuracy and stability.

[0059] The character recognition model is constructed using a convolutional neural network structure, as follows: The first layer is a convolutional layer with a kernel size of 3 pixels × 3 pixels, 32 channels, a stride of 1, and a rectified linear unit function as the activation function. The second layer is a max pooling layer with a pooling window size of 2 pixels × 2 pixels and a stride of 2. The third layer is a convolutional layer with a kernel size of 3 pixels × 3 pixels and 64 channels. The fourth layer is the max pooling layer, with the same structure as above; The 5th layer is a fully connected layer with an output feature dimension of 128; The 6th layer is the output layer, with the number of nodes equal to the total number of character categories (e.g., 36 categories, including digits 0 to 9 and uppercase letters A to Z). The Softmax function is used to output the probability distribution.

[0060] The model is trained under supervision on a dedicated training dataset containing character images of tire molds. The dataset contains over 50,000 images, covering different fonts, character depths, mold materials, and lighting conditions. Training employs the cross-entropy loss function and the Adam optimizer, with an initial learning rate of 0.0005, 100 training epochs, and a batch size of 64. After training, forward inference is performed on the input character images, outputting a character category probability vector. The character with the highest probability is selected as the recognized character, and its corresponding recognition confidence value is recorded, ranging from 0 to 1.

[0061] After the recognition result is generated, the production scheduling data interface is called to obtain the mold character template data corresponding to the current detection time point. This template data is a set of target characters arranged in sequence, which usually includes tire model code, batch identifier, date code, etc., and the total number of characters is consistent with the number of mold character areas.

[0062] The identified character categories are sorted according to their image coordinates and compared one-to-one with the template character sequence. The comparison method is exact matching. If a character does not match the template character, or if the confidence score of the character is lower than a set threshold (e.g., 0.85), it is marked as a mismatched character item.

[0063] After completing the character-by-character comparison, the number and position of mismatched characters are counted, and the character matching results are generated. If all characters match and the confidence level is not lower than the preset threshold, the character recognition result of the template is marked as "match passed"; if there is any mismatch or insufficient confidence level, it is marked as "match abnormal", and the position and content of the abnormal character are recorded.

[0064] The matching result will serve as an automatic criterion for determining whether the mold installation conforms to the current production schedule. It will also be stored in the inspection dataset as part of the inspection data for subsequent statistical analysis and quality traceability.

[0065] Through the above character recognition and matching process, mold character verification can be completed automatically without human intervention, effectively improving the efficiency and accuracy of mold installation on the production line and significantly reducing the risk of human error.

[0066] If there are character recognition results that do not match the template or whose confidence level is lower than the set threshold, the character area is marked as an abnormal area and a corresponding alarm message is generated.

[0067] After recognizing the character image and comparing it with the template character, to ensure that the mold character identification meets the current production schedule requirements, it is necessary to judge the recognition results, identify potentially abnormal character areas, and generate traceable alarm outputs. This implementation method achieves automatic verification of character recognition accuracy and an anomaly alert mechanism by setting a recognition confidence threshold and comparison result judgment logic. The specific steps are as follows: After character recognition is completed, the system obtains the character template sequence that the current mold should contain. Each character recognition result consists of two parts: Identify character categories, that is, identify the character with the highest probability output by the model (such as "A", "7", "T" etc.). The confidence value is identified, which is the maximum probability value output by Softmax, ranging from 0 to 1.

[0068] The identified characters are compared one-to-one with the template character sequence according to their arrangement in the image. The following judgment logic is performed for each character position: If the recognized characters do not match the template characters, it is judged as "content mismatch"; If the identified character matches the template character, but its confidence level is lower than the set threshold, it is also judged as "insufficient confidence level".

[0069] The confidence threshold is a system preset parameter, preferably set to 0.85. This value was determined based on extensive experiments and can ensure recognition accuracy while avoiding false alarms due to oversensitivity.

[0070] If any character position is detected to have one of the two situations mentioned above, namely "content mismatch" or "insufficient confidence", then the image region corresponding to that character is marked as an abnormal region.

[0071] The labeling information for abnormal areas includes the following fields: the image coordinate range of the abnormal character (top-left horizontal and vertical coordinates, width, and height); the recognized character; the corresponding character in the template; the actual confidence score; and the anomaly type (content mismatch / insufficient confidence). This labeling information will be appended to the character recognition results and highlighted in the visualization interface to help operators quickly locate the problem area.

[0072] When at least one abnormal character region is detected in the test results, the system will trigger an alarm mechanism and generate a complete alarm message record. The alarm message includes, but is not limited to, the following: Mold number and inspection timestamp; abnormal character location and corresponding content; alarm level (can be divided into prompt / serious error, based on the number of abnormalities); suggested operation (such as "Please check the mold installation direction" or "Please replace the mold component").

[0073] Alarm information is output in the following ways: pushed to the display terminal interface of the mold installation station; recorded in the database for subsequent statistics and traceability; and optionally pushed to quality management personnel via email or SMS.

[0074] This anomaly identification and alarm mechanism enables automatic verification and feedback of character recognition quality, ensuring that mold markings conform to the production schedule and avoiding quality problems such as mold assembly errors or mislabeling of product information caused by character errors, thus significantly improving the intelligence and reliability of the tire mold installation process.

[0075] The alarm information, along with the original image and character position coordinates, is stored in the mold detection database. The abnormal detection results are then fed back to the mold installation terminal through the MES system to provide on-site abnormal alerts.

[0076] After character recognition and abnormal alarm information generation, this implementation provides a complete data writing and feedback mechanism to enable tracking, recording, and on-site intervention control of the detection results. By jointly storing the detection image, character positioning data, and alarm information in the mold detection database, and simultaneously pushing the abnormal detection results to the mold installation station display terminal, information closure and timely alerts are achieved. The specific steps are as follows: When a character recognition anomaly is detected, the system generates a complete alarm message record. To ensure structured data management and traceability, the alarm message is stored together with the original test data in the mold testing database.

[0077] Each record includes the following fields: Mold identification information: includes mold number, inspection timestamp, and inspection station number; Raw image data: Images of the mold area captured by an industrial camera and processed uniformly, compressed and stored in JPEG format; Character position coordinates: including the horizontal and vertical coordinates of the top left corner of all candidate character regions, as well as their width and height, used for image relocation; Abnormal character information: including the recognition result, template character, recognition confidence, and abnormality type for each abnormal character; Alarm level: Automatically assessed based on the number and type of abnormal characters, categorized into three levels: "Prompt", "Warning", or "Critical". Processing status: Used to record the processing results after manual confirmation or re-inspection (e.g., "Confirmed" or "Mold replaced").

[0078] The database adopts a relational structure design and preferably uses a structured query language compatible database system that supports large-scale image storage and fast querying. The character coordinate field and image records are linked by an index through foreign keys to support image positioning and visualization playback functions.

[0079] To provide real-time alerts for abnormal detection during mold installation, the detection results are synchronously pushed to the terminal display device at the current workstation. The terminal device establishes a data interface connection with the Manufacturing Execution System (MES) to automatically send out abnormal results, specifically including: Push notification trigger timing: If an anomaly is detected within 500 milliseconds after the character recognition and comparison process is completed, a push notification will be triggered immediately. Push notification content format: includes mold number, abnormal character location information, abnormal category, recommended handling suggestions, and a thumbnail image of the abnormal character area; Display method: The workstation terminal uses a combination of text and graphics to display the abnormal content, highlighting the abnormal characters with a red box, and providing specific template comparison failure information and confidence level values. Workstation response operations: The terminal interface supports interactive operations such as operators confirming anomalies, requesting re-inspection, and uploading processing opinions. All operation records are synchronously written back to the database.

[0080] The manufacturing execution system (MES) uses a RESTful interface for data integration, enabling standardized information transmission based on the HTTP protocol. Each anomaly detection result is encapsulated in JSON format, containing complete fields. The interface response time is preferably controlled within 100 milliseconds to ensure real-time synchronization between anomaly alerts and detection operations.

[0081] Through the above processing methods, the mold inspection system can not only automatically determine character recognition anomalies, but also has a complete recording, feedback and human interaction mechanism, forming a closed loop of detection-prompt-traceability, effectively ensuring the accuracy of the tire mold installation process and the real-time quality control.

[0082] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An automatic character detection method for tire molds based on image processing, characterized in that: include: The image of the tire mold is acquired by an industrial camera located at the mold inspection station under fixed lighting conditions. The mold includes at least one area for engraving characters on a curved surface. The mold image is segmented into regions, and a candidate region is generated for regions in the image that may contain characters based on a deep convolutional feature extraction network, resulting in several candidate character regions. Geometric morphology correction is performed on each candidate character region. The character arrangement direction is determined based on the character edge density and curvature characteristics. Direction normalization is then performed to obtain the character forward image. Input the forward image of the character into the pre-trained character recognition model to identify the content of each character and compare it with the corresponding mold character template data in the current production schedule to generate character matching results; If there are character items whose character recognition results do not match the template or whose confidence level is lower than the set threshold, the character area is marked as an abnormal area and a corresponding alarm message is generated. The alarm information, along with the original image and character position coordinates, is stored in the mold detection database. The abnormal detection results are then fed back to the mold installation terminal through the MES system to provide on-site abnormal alerts.

2. The automatic character detection method for tire molds based on image processing according to claim 1, characterized in that: The step of generating candidate character regions includes: The mold image is subjected to grayscale normalization and Gaussian filtering to enhance the contrast between character edges and background; Based on the image pyramid structure, a sliding window scan is performed on multi-scale images to initially screen out image sub-regions with character texture features; Feature vectors are extracted from image sub-regions using a feature extraction model built on a deep convolutional neural network, and then input into a classifier for character region discrimination. The coordinates of the image sub-regions identified as character regions are used as candidate character regions.

3. The automatic character detection method for tire molds based on image processing according to claim 1, characterized in that: The step of performing geometric morphological correction on each candidate character region includes: Edge extraction processing is performed on the candidate character region. The Canny algorithm is used to extract the character outline and calculate the edge density distribution of the outline pixels in the horizontal and vertical directions. The main direction of the character is determined based on the direction of maximum edge density, and the angle of the main axis of the character arrangement is extracted by combining the minimum bounding rectangle fitted by the character outline within the region. Affine transformation correction is performed on the character region image based on the principal axis direction angle to align the character arrangement direction with the horizontal direction of the image. The corrected image is then processed for edge completion and size normalization to generate a character-oriented image region.

4. The automatic character detection method for tire molds based on image processing according to claim 3, characterized in that: The step of calculating the edge density distribution of contour pixels in the horizontal and vertical directions includes: The extracted contour image is scanned by row and column respectively, and the number of edge pixels in each row and column is counted. Divide the number of edge pixels in each row by the total number of pixels in that row to obtain the horizontal edge density sequence; divide the number of edge pixels in each column by the total number of pixels in that column to obtain the vertical edge density sequence. The mean and variance of the edge density sequences in the horizontal and vertical directions are calculated respectively to measure the central tendency and directional consistency of the edge distribution. By comparing the average density and variance of the two directions, the main direction of the edge distribution is determined, which helps to determine the main axis angle of the character arrangement direction.

5. The automatic character detection method for tire molds based on image processing according to claim 3, characterized in that: The step of extracting the angle of the main axis of the character arrangement includes: By comparing the average values ​​of the horizontal and vertical edge density sequences in the candidate character region, the main direction of the edge density distribution is determined. Based on the determined main direction, the coordinate set of the pixel points of the character region outline is fitted, and the least squares method is used to fit its minimum bounding rectangle. Extract the angle between the long side of the minimum bounding rectangle and the horizontal direction of the image, and use it as the angle of the main axis direction of character arrangement; If the spindle direction angle exceeds the preset angle threshold range, the direction normalization process is triggered.

6. The automatic character detection method for tire molds based on image processing according to claim 1, characterized in that: The steps for generating character matching results include: The forward-facing character images that have undergone orientation and size normalization are input into the pre-trained character recognition model with uniform pixel resolution and grayscale range. A character recognition model based on a convolutional neural network is used to extract features and classify character images, and output the corresponding character category and its recognition confidence. Obtain the mold character template data corresponding to the current production schedule, and compare the character categories with the target characters in the template data one by one; Based on the consistency between the character recognition result and the template character, and whether the recognition confidence level meets the preset threshold conditions, a character matching result is generated.

7. The automatic character detection method for tire molds based on image processing according to claim 1, characterized in that: The industrial camera is mounted on a fixed bracket with multi-degree-of-freedom adjustment capability, and is illuminated by a ring-shaped light-emitting diode light source within a closed or semi-closed light cover, so that the curved surface engraved character area forms a stable light and dark contrast feature in the image.

8. The automatic character detection method for tire molds based on image processing according to claim 6, characterized in that: The generation of the character matching result includes: comparing the character recognition result with the corresponding mold character template data in the production schedule according to the spatial order of the character region in the mold image.

9. The automatic character detection method for tire molds based on image processing according to claim 8, characterized in that: When the character recognition result is inconsistent with the template character, or the character recognition confidence is lower than the set threshold, the corresponding character area is marked as an abnormal area, and the abnormal type and abnormal location coordinates are recorded.

10. The automatic character detection method for tire molds based on image processing according to claim 3, characterized in that: After determining the main direction of the character, the minimum bounding rectangle is fitted to the set of character outline pixels, and the angle between the long side of the minimum bounding rectangle and the horizontal direction of the image is extracted as the angle of the main axis of the character arrangement. When the angle of the main axis of the character arrangement exceeds the preset angle threshold, affine transformation correction is performed on the candidate character region to align the character arrangement direction with the horizontal direction of the image.

Images