A low false rejection rate detection method and system for tire surface defects
Patent Information
- Application Number
- CN202611218308.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-10-23
AI Technical Summary
[0024]本发明是为了解决上述问题而进行的，目的在于提出一种适用于轮胎表面缺陷检测的低漏检率检测方法，通过样本增强、目标检测粗定位、语义分割精判定以及后处理判决策略协同工作，解决现有技术中轮胎表面小缺陷漏检率高、复杂纹理误检多、过杀率高以及缺陷样本不足导致泛化能力差的问题
[0035]与现有技术相比，采用本发明提供的面向轮胎表面缺陷的低漏检率检测方法及系统，可以实现如下技术效果：
Smart Images
Figure CN122966618A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial visual inspection and artificial intelligence technology, specifically relating to a method and system for detecting tire surface defects with a low missed detection rate. Background Technology
[0002] As a critical safety component in vehicle operation, the surface quality of tires directly affects vehicle safety, durability, and product yield. During tire production, factors such as raw materials, molding, vulcanization, transportation, storage, and manual handling can lead to hundreds of defects on the tire surface, including cracks, scratches, bubbles, bulges, missing rubber, foreign objects, indentations, contamination, and abnormal tread patterns. These defects are often small in size, have indistinct boundaries, and irregular shapes. Furthermore, tire surfaces typically exhibit complex textures, periodic patterns, curved surfaces, and reflective variations, making defect detection highly challenging.
[0003] Traditional tire surface inspection mainly relies on manual visual inspection or traditional machine vision algorithms. Manual inspection involves holding a lamp and inspecting each part of the tire, first the inside and then the outside, while the tire is rotating. This process is subject to problems such as high subjectivity, low efficiency, fatigue leading to missed detections, and inconsistent inspection standards. Traditional machine vision methods usually rely on rules such as grayscale thresholding, edge detection, template matching, and texture analysis. They are sensitive to changes in lighting, angle, tire model, and surface texture, making them difficult to adapt to complex and ever-changing industrial environments.
[0004] In recent years, deep learning methods have been increasingly applied to the field of industrial defect detection. Common approaches include defect localization based on object detection models, defect region extraction based on semantic segmentation models, OK / NG determination based on classification models, and unsupervised defect detection based on anomaly detection models. Among these, the YOLO series of object detection models offer advantages such as fast detection speed, convenient deployment, and suitability for real-time industrial inspection; semantic segmentation models such as Unet, DeepLabV3+, and SegFormer possess pixel-level region extraction capabilities, enabling more precise descriptions of defect boundaries.
[0005] However, in tire surface defect detection scenarios, single-target detection models are easily affected by small targets, low-contrast defects, and complex textured backgrounds, leading to missed detections. If the target detection confidence threshold is directly lowered to reduce the missed detection rate, a large number of suspected defect boxes will be generated, increasing the false detection rate and over-detection rate. Although single segmentation models can obtain relatively fine defect contours, directly segmenting the entire large image is easily affected by background texture interference, has a large computational load, and lacks stability for small defects.
[0006] Due to the vast number of tire models and the diverse tread patterns and styles, conventional image processing and machine learning methods are often ineffective. However, the rapid development of AI in recent years has led to new technological solutions, typically employing the following AI technical approaches:
[0007] The first approach is to directly utilize an object detection model for tire surface defect detection. This approach first acquires tire surface images, labels defect areas with rectangular bounding boxes, and then trains the object detection model. During the inference phase, the image to be detected is input into the object detection model, and the model outputs the defect category, defect location bounding box, and confidence score. When the confidence score is greater than a set threshold, the area is determined to have a defect. The basic process is: tire image acquisition → defect bounding box labeling → object detection model training → defect bounding box output → confidence threshold determination.
[0008] The second approach is to directly use a semantic segmentation model for defect detection. This approach involves pixel-level annotation of defect regions. The entire tire image or a local image is input into the segmentation model, which outputs a defect mask. The presence of a defect is then determined based on the mask's area, shape, or category. The basic workflow is: tire image acquisition → pixel-level defect annotation → segmentation model training → defect mask output → defect determination.
[0009] The third approach is to use conventional data augmentation during the training phase. Common augmentation methods include image rotation, flipping, scaling, cropping, brightness adjustment, contrast adjustment, noise addition, blurring, Mosaic, MixUp, Copy-Paste, etc., to expand the number of training samples and improve the model's generalization ability.
[0010] The fourth approach is to combine a detection model with a classification or segmentation model. The target detection model first outputs candidate defect boxes, and then the images within the candidate boxes are classified or segmented for verification to reduce false alarms.
[0011] Currently, existing technologies have the following shortcomings:
[0012] 1. Small target defects have a high rate of missed detection.
[0013] Tire surface defects include numerous small cracks, shallow scratches, pinholes, localized contamination, and minor rubber defects. These defects are small in size and have low contrast, making them prone to feature loss after image scaling, and single-target detection models struggle to achieve stable recall.
[0014] 2. Complex textures lead to false positives and over-detection.
[0015] Tire surfaces feature grooves, patterns, characters, tire hairs, reflective areas, and textural variations. Traditional object detection models are prone to misclassifying normal textures as defects, especially when lowering the confidence threshold to increase recall, which can lead to a large number of false positives and an increased over-detection rate.
[0016] 3. The single confidence level method is unreliable.
[0017] Existing object detection schemes typically rely on the confidence score of the detection bounding box for final determination, but the confidence score cannot accurately reflect the true shape, area, boundary, and connectivity of the defect. For low-contrast defects, the model confidence score may be too low; for complex textures, the model confidence score may be artificially high.
[0018] 4. Direct whole-image segmentation involves large computational costs and is easily affected by background interference.
[0019] If a semantic segmentation model is applied directly to the entire tire image, it needs to process a large amount of irrelevant background textures, resulting in high inference overhead. Furthermore, it is easily affected by tire tread patterns, reflections, and character regions, leading to unstable segmentation results.
[0020] 5. Insufficient defective samples limit the model's generalization ability.
[0021] In industrial settings, real-world defect samples are often few in number and unevenly distributed, with some severe defects being difficult to collect. Relying solely on conventional data augmentation is insufficient to generate a sufficiently rich variety of defect morphologies, and the model's performance is prone to degradation under new tire models, new lighting conditions, or new defect morphologies.
[0022] 6. Lack of effective coordination between detection and segmentation
[0023] Some existing two-stage solutions simply connect the detection model and the segmentation model in series. They do not perform low-threshold candidate generation, candidate box expansion, ROI context preservation, segmentation verification, and morphological rule determination for high-recall targets. Therefore, it is difficult to simultaneously achieve low false negative rate and low false positive rate. Summary of the Invention
[0024] This invention is made to solve the above-mentioned problems. The purpose is to propose a low-miss-rate detection method for tire surface defect detection. By working together with sample enhancement, coarse target detection localization, semantic segmentation fine judgment, and post-processing decision strategy, it solves the problems of high miss rate of small tire surface defects, many false detections of complex textures, high over-detection rate, and poor generalization ability due to insufficient defect samples in the prior art.
[0025] This invention provides a low-miss-rate detection method for tire surface defects, comprising the following steps: Step S1, image acquisition and defect annotation: acquiring a tire surface image and annotating the defect regions in the image with target detection boxes and pixel-level segmentation masks; Step S2, sample augmentation processing: performing sample augmentation processing on the annotated training samples to generate augmented training samples with diverse forms; Step S3, model training and candidate box generation: inputting the augmented training samples into the target detection model for training; during the detection phase, the tire surface image to be detected is input into the target detection model with a threshold value lower than a preset standard threshold. Step S4, ROI Expansion and Segmentation Confirmation: The suspected defect candidate boxes are expanded, and ROIs that retain the surrounding context information of the defect are cropped. The ROIs are then input into the semantic segmentation model to output a pixel-level segmentation mask for the defect. Step S5, Comprehensive Defect Judgment: Based on the morphological features and segmentation confidence of the pixel-level segmentation mask, the suspected defect candidate boxes are comprehensively judged, and the final defect detection result is output. Step S6, Result Output and Closed-Loop Management: The final defect detection result is output, and quality statistics and closed-loop data management are performed based on the final defect detection result.
[0026] Preferably, in step S2, the sample enhancement process is based on the O2MAG sample diffusion enhancement algorithm, which is used to generate defect samples with diverse shapes, scales, texture fusion effects and background adaptability.
[0027] Preferably, in step S3, the target detection model includes a small target recognition enhancement module and a GA geometric attention module; the small target recognition enhancement module is used to strengthen the shallow fine-grained feature representation; the GA geometric attention module is used to guide the model to focus on the geometric morphological features of tire surface defects.
[0028] Preferably, in step S4, the expansion process is as follows: based on the width w, height h and expansion ratio r of the suspected defect candidate box, the expansion is performed. The width w' = w × (1 + r) and the height h' = h × (1 + r) of the expanded candidate region are, where the value of r ranges from 0.1 to 0.3.
[0029] Preferably, in step S4, the semantic segmentation model is a DeepLabV3+ variant segmentation model, which includes a GA geometric attention module to enhance the model's pixel-level recognition ability for defect boundaries, slender structures and irregular shapes.
[0030] Preferably, in step S5, the morphological features include the defect mask area, the number of connected components, the aspect ratio, the contour shape, the defect length, and the defect width; the comprehensive judgment includes: when the defect mask area is greater than a preset area threshold, the maximum connected component meets the preset morphological conditions, and the segmentation confidence is greater than a preset confidence threshold, the current suspected defect candidate box is determined to be a real defect.
[0031] Preferably, in step S5, the final defect detection result includes: defect category, defect location, defect pixel-level segmentation mask, defect area, defect length, defect width, defect confidence level, and product OK / NG determination result.
[0032] Preferably, in step S6, the final defect detection result is output to a display terminal, host computer, PLC, MES or quality management system, thereby triggering an alarm, rejection, re-inspection or manual review process for the defective tire.
[0033] This invention provides a low-miss-rate detection system for tire surface defects, comprising: an image acquisition and defect annotation module for acquiring tire surface images and annotating defect regions in the images with target detection bounding boxes and pixel-level segmentation masks; a sample enhancement processing module for enhancing the annotated training samples to generate enhanced training samples with diverse morphologies; a model training and candidate box generation module for inputting the enhanced training samples into a target detection model for training; during the detection phase, the tire surface image to be detected is input into the target detection model, and suspected defect candidate boxes are output with a low confidence threshold below a preset standard threshold; a ROI expansion and segmentation confirmation module for expanding the suspected defect candidate boxes, cropping out ROIs that retain the surrounding context information of the defects, and inputting the ROIs into a semantic segmentation model to output a defect pixel-level segmentation mask; a defect comprehensive judgment module for comprehensively judging the suspected defect candidate boxes based on the morphological features and segmentation confidence of the defect pixel-level segmentation mask, and outputting the final defect detection result; and a result output and closed-loop management module for outputting the final defect detection result and performing quality statistics and closed-loop data management based on the final defect detection result.
[0034] Technical effect
[0035] Compared with existing technologies, the low-miss-rate detection method and system for tire surface defects provided by this invention can achieve the following technical effects:
[0036] This invention introduces a sample diffusion enhancement mechanism. Based on conventional data augmentation, it further utilizes the sample diffusion enhancement method to generate diverse defect samples, thereby enhancing the model's ability to identify defects under different shapes, scales, locations, textures, backgrounds, and lighting conditions, and alleviating the problem of insufficient real defect samples.
[0037] This invention constructs a coarse localization mechanism for low-confidence, high-recall target detection. The target detection model does not prioritize high-precision final determination, but rather aims for high candidate recall. By lowering the detection confidence threshold, enhancing small target features, and introducing a geometric attention mechanism (GA), more suspected defect regions are preserved, thereby reducing the probability of missed detections in the first stage.
[0038] This invention employs a detection-segmentation cascade structure. In the first stage, a target detection model is responsible for quickly locating suspected defect areas. In the second stage, an optimized segmentation model performs pixel-level fine segmentation of the candidate areas. This structure combines "high recall and coarse localization" with "fine segmentation verification," avoiding the limitations of a single detection model or a single segmentation model.
[0039] This invention expands the candidate bounding box while preserving contextual information. Before sending the detection box into the segmentation model, the candidate bounding box is expanded outward by a certain proportion, so that the defect edge and surrounding texture context are preserved, avoiding the defect area being truncated due to the detection box being too tight, thus improving the segmentation stability.
[0040] This invention is based on multi-condition defect determination of segmented masks. The final determination no longer depends solely on the target detection confidence, but combines features such as segmented mask area, number of connected components, aspect ratio, contour shape, segmentation confidence, defect length / width, etc. to make a comprehensive judgment, thereby filtering out false defects caused by complex textures and reducing the over-detection rate.
[0041] This invention focuses on overall optimization to achieve a low false negative rate. Through a collaborative approach of "sample enhancement to improve generalization ability, detection stage to improve candidate recall, segmentation stage to reduce false alarms, and post-processing stage to control overkill", this invention enables the system to achieve a defect detection rate of over 95%, a false negative rate of close to 1%, and an overkill rate of less than 10% in tire surface defect detection. Attached Figure Description
[0042] Figure 1 This is a flowchart of a low-miss-rate detection method for tire surface defects in an embodiment of the present invention. Detailed Implementation
[0043] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments are described in detail with reference to the accompanying drawings.
[0044] This embodiment provides a low-miss-rate detection method for tire surface defects.
[0045] Figure 1 This is a flowchart of a low-miss-rate detection method for tire surface defects in an embodiment of the present invention.
[0046] like Figure 1As shown, the low-miss-rate detection method for tire surface defects in this embodiment includes the following steps:
[0047] Step S1, Image Acquisition and Defect Annotation: Acquire an image of the tire surface and perform target detection bounding box annotation and pixel-level segmentation mask annotation on the defect areas in the image. Specifically, this includes the following steps:
[0048] Step S1-1, Tire Surface Image Acquisition: Tire surface images are acquired using a hybrid mode of industrial cameras, line scan cameras, or area scan cameras. The acquired areas include the tread, sidewall, shoulder, crown, toe, and inner sidewall. During acquisition, light source control, trigger acquisition, tire rotation mechanisms, or conveying mechanisms can be combined to obtain continuous, clear, and stable tire surface images. The acquired images are stored and managed according to information such as tire model, acquisition location, batch, and defect type.
[0049] Step S1-2, Defect Sample Annotation: Annotate the defects in the acquired images. Annotation methods include:
[0050] (1) Object detection box annotation: used to train the first-stage object detection model.
[0051] (2) Pixel-level mask annotation: used to train the second-stage segmentation model.
[0052] (3) Defect category labeling: used to distinguish defect types such as cracks, scratches, missing glue, bulges, foreign objects, and contamination.
[0053] (4) OK / NG image level annotation: used for product level judgment and statistics.
[0054] (5) The same defect sample can have both a detection box and a pixel-level mask annotation.
[0055] Step S2, Sample Augmentation: The labeled training samples are augmented to generate augmented training samples with diverse formats. Augmentation methods include two categories:
[0056] The first category is conventional data augmentation, including but not limited to: rotation, flipping, scaling, random cropping, brightness adjustment, contrast adjustment, color perturbation, blurring, noise, Mosaic, MixUp, and Copy-Paste.
[0057] The second type is O2MAG sample diffusion enhancement. This enhancement method is used to generate defect samples with diverse shapes, scales, texture blending effects, and background adaptability. Through O2MAG sample diffusion enhancement, the number of samples of low-frequency defects, small target defects, and blurred boundary defects can be increased, thereby improving the model's generalization ability to complex defect morphologies.
[0058] The enhanced samples and the original samples together form the training dataset.
[0059] Step S3, Model Training and Candidate Box Generation: Enhanced training samples are input into the target detection model for training. During the detection phase, the tire surface image to be detected is input into the target detection model, and suspected defect candidate boxes are output with a low confidence threshold below a preset standard threshold. The specific process is as follows:
[0060] Step S3-1, in this embodiment, the first stage of target detection model training: constructing a target detection model, preferably a target detection variant model. This model includes a small target recognition enhancement module and a GA geometric attention module.
[0061] The small target recognition enhancement module is used to strengthen the expression of shallow fine-grained features and improve the response capability to micro-cracks, shallow scratches and local abnormal textures.
[0062] The GA geometric attention module guides the model to focus on the geometric features of tire surface defects, such as slender structures, abrupt edge changes, local bending, and abnormal texture orientation, thereby improving the model's ability to locate defect areas.
[0063] After training, the first-stage object detection model is used to output candidate boxes for suspected defects.
[0064] Step S3-2, Generation of Low-Confidence, High-Recall Candidate Boxes: During the inference phase, the tire image to be detected is input into the first-stage target detection model. Unlike conventional target detection methods, this embodiment uses a lower confidence threshold to retain more potential defect candidate boxes. The conventional confidence threshold is 0.3-0.5, while this embodiment uses 0.1-0.2.
[0065] To address the slow inference speed of high-resolution images, this embodiment employs an image segmentation strategy: a sliding 640*640 fixed-size box is used, which slides at a certain step size. Each slide segments an image. The overlap rate between each image can be set from 5% to 20%. The overlap is generally selected as an integer multiple of 32. In this embodiment, 128 is used as the overlap value.
[0066] In this embodiment, the target detection model outputs a set of candidate boxes: B = {b1, b2, ..., bn}.
[0067] Each candidate box includes location coordinates, defect category, and confidence level. By lowering the confidence threshold, regions with low confidence but potentially real defects are not discarded in the first stage, thereby improving candidate recall.
[0068] Meanwhile, the NMS threshold and the maximum number of candidate boxes can be adjusted to avoid excessive suppression of adjacent small defects.
[0069] Step S4, ROI Expansion and Segmentation Confirmation: The suspected defect candidate boxes are expanded to extract the ROIs that retain the surrounding context information of the defect. These ROIs are then input into the semantic segmentation model, which outputs a pixel-level defect segmentation mask. The specific process is as follows:
[0070] Step S4-1, Candidate Box Expansion and ROI Clipping:
[0071] The candidate bounding boxes output in the first stage are expanded. Let the width of the candidate bounding box be w, the height be h, and the expansion ratio be r. Then the expanded candidate region can be represented as:
[0072] w' = w × (1 + r)
[0073] h' = h × (1 + r)
[0074] The value of r is set according to the defect scale and image resolution, and can be selected from 0.1 to 0.3. The larger the defect size, the smaller the value of the expansion ratio r. For example, if the minimum bounding rectangle of the defect size is 5*8mm, then choosing r=0.1 will result in a candidate box width of 5.5mm and a height of 8.8mm. After expanding the candidate box, the ROI region is cropped from the original image. This operation can preserve the context around the defect, avoid the defect edges being truncated, and improve the fine recognition ability of subsequent segmentation models.
[0075] Step S4-2, second-stage segmentation model fine segmentation:
[0076] The cropped ROI is input into the DeepLabV3+ variant segmentation model. This segmentation model can incorporate a GA geometric attention module to enhance the model's pixel-level recognition capabilities for defect boundaries, slender structures, and irregular shapes.
[0077] The segmentation model outputs the defect mask and segmentation confidence for each ROI. The defect mask represents the pixel-level region of the defect.
[0078] The purpose of this stage is to refine the suspected defect areas output in the first stage and filter out false defects caused by normal tire texture, characters, reflections, tread edges, etc.
[0079] Step S5, Defect Comprehensive Judgment: Based on the morphological features and segmentation confidence of the defect pixel-level segmentation mask, a comprehensive judgment is made on the suspected defect candidate boxes, and the final defect detection result is output. The specific process is as follows:
[0080] Post-processing of the mask output by the segmentation model, including but not limited to:
[0081] (1) Remove small connected components;
[0082] (2) Morphological opening and closing operations;
[0083] (3) Smooth boundaries;
[0084] (4) Connectivity analysis;
[0085] (5) Area calculation;
[0086] (6) Length and width calculation;
[0087] (7) Calculation of aspect ratio;
[0088] (8) Calculation of the proportion of defects to ROI;
[0089] (9) Statistical analysis of split confidence.
[0090] Based on the above characteristics, a final comprehensive judgment is made, using multi-condition OR, AND, and other logic to output the judgment. In this embodiment, "AND" logic is used for judgment. When the defect mask area is greater than a preset area threshold (100 pixels in this embodiment), the largest connected component meets the morphological condition, the segmentation confidence is greater than the threshold of 0.9, and the defect shape (slender, circular, rectangular, etc.) conforms to the corresponding category rules, the candidate region is determined to be a real defect; otherwise, it is determined to be a false defect.
[0091] The final output includes:
[0092] (1) Defect category;
[0093] (2) Location of the defect;
[0094] (3) Defect mask;
[0095] (4) Defect area;
[0096] (5) Defect length and width;
[0097] (6) Confidence level of defects;
[0098] (7) Product OK / NG judgment result.
[0099] Step S6, Result Output and Closed-Loop Management: Output the final defect detection results and perform quality statistics and closed-loop data management based on the final defect detection results.
[0100] In this embodiment, the system outputs the final inspection results to a display terminal, host computer, PLC, MES, or quality management system. For tires with defects, an alarm, rejection, re-inspection, or manual review process can be triggered.
[0101] The system can also collect the following metrics:
[0102] (1) Defect detection rate;
[0103] (2) False negative rate;
[0104] (3) Overkill rate;
[0105] (4) False detection rate;
[0106] (5) Recall rates for different defect categories;
[0107] (6) Performance testing for different tire models;
[0108] (7) Model inference time.
[0109] By continuously collecting false positive and false negative samples, a closed-loop dataset can be formed for subsequent model iteration and optimization.
[0110] This embodiment also provides a low-miss-rate detection system for tire surface defects, including:
[0111] The image acquisition and defect annotation module is used to execute the process in step S1 of this embodiment.
[0112] The sample enhancement processing module is used to execute the process in step S2 of this embodiment.
[0113] The model training and candidate box generation module is used to execute the process in step S3 of this embodiment.
[0114] The ROI expansion and segmentation confirmation module is used to execute the process in step S4 of this embodiment.
[0115] The defect comprehensive judgment module is used to execute the process in step S5 of this embodiment.
[0116] The result output and closed-loop management module is used to execute the process in step S6 of this embodiment.
[0117] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.
[0118] This embodiment enhances the diversity of defect samples through sample diffusion enhancement, improves candidate recall through low-confidence target detection, refines candidate regions through a segmentation model, and makes a final judgment based on the morphological features of the segmentation mask. This approach effectively reduces the false negative rate and controls the false negative rate in complex tire surface texture environments, thereby improving the defect detection rate. This method is particularly suitable for tire surface defect detection scenarios involving small targets, low contrast, blurred boundaries, and irregular shapes.
[0119] In the scenario of tire surface defect detection, this embodiment can reduce the missed detection rate and the over-detection rate while maintaining a high defect detection rate. Under experimental conditions, the defect detection rate can reach more than 95%, the missed detection rate can be reduced to about 1%, and the over-detection rate can be controlled below 10%.
[0120] This embodiment has the following beneficial effects:
[0121] (1) Low false negative rate and low false negative rate: By using AI technology, the problem of high false negative rate is greatly reduced, the probability of defective products flowing to the next process is reduced, the problem of customer quality complaints caused by false negatives is improved, and the false negative rate is greatly reduced, greatly reducing the amount of manual secondary inspection and judgment.
[0122] (2) High detection accuracy: The use of high-resolution laser line scanning solves the problem of high measurement accuracy and eliminates the large errors caused by relying on visual measurement with tools, thus improving the evaluation level of detection quality. The height measurement accuracy of this scheme can be controlled at around 10μm, while the measurement accuracy of the previous manual detection scheme was 0.2mm, representing a 20-fold improvement in accuracy.
[0123] (3) Wide detection range: Multiple parallel scans using a high-resolution laser line scanner are stitched together to solve the problem of product coverage. For products with a large area, this method can be used to obtain a complete high-definition scan image of the product.
[0124] (4) Fast detection speed: solves the problem of slow inference speed at high resolution. The detection uses a cascaded model to solve the problem of rapid localization and segmentation of product features in high-resolution images. For the high-resolution images in the first stage, this embodiment uses the image segmentation function to divide the original image into multiple image sizes suitable for the model, which effectively ensures the detection accuracy of the original high-resolution image. At the same time, the defective images are subjected to ROI operation, and the small-sized ROI images are fed to the segmentation model for high-precision segmentation, which greatly reduces the inference complexity of segmentation and ensures the production line's generation cycle time.
[0125] (5) It has promoted the possibility of this type of detection from manual detection to fully automated detection, replacing a large number of manual workers, reducing detection costs, and improving detection efficiency and quality.
Claims
1. A method for detecting tire surface defects with a low false negative rate, characterized in that, Includes the following steps: Step S1, Image Acquisition and Defect Annotation: Acquire images of the tire surface and annotate the defect areas in the images with target detection boxes and pixel-level segmentation masks; Step S2, Sample Augmentation Processing: Perform sample augmentation processing on the labeled training samples to generate augmented training samples with diverse forms; Step S3, Model Training and Candidate Box Generation: Input the enhanced training samples into the object detection model for training; During the detection phase, the image of the tire surface to be detected is input into the target detection model, and a candidate box for suspected defects is output with a low confidence threshold lower than the preset standard threshold. Step S4, ROI expansion and segmentation confirmation: The suspected defect candidate box is expanded to extract the ROI that retains the surrounding context information of the defect, and the ROI is input into the semantic segmentation model to output the defect pixel-level segmentation mask; Step S5, Defect Comprehensive Judgment: Based on the morphological features and segmentation confidence of the defect pixel-level segmentation mask, the suspected defect candidate box is comprehensively judged, and the final defect detection result is output. Step S6, Result Output and Closed-Loop Management: Output the final defect detection result and perform quality statistics and closed-loop data management based on the final defect detection result.
2. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S2, the sample enhancement process is based on the O2MAG sample diffusion enhancement algorithm, which is used to generate defect samples with diverse shapes, scales, texture fusion effects and background adaptability.
3. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S3, the target detection model includes a small target recognition enhancement module and a GA geometric attention module; The small target recognition enhancement module is used to strengthen the expression of shallow, fine-grained features; The GA geometric attention module is used to guide the model to focus on the geometric features of tire surface defects.
4. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S4, the expansion process is as follows: based on the width w, height h and expansion ratio r of the suspected defect candidate box, the expansion is performed. The width w' = w × (1 + r) and the height h' = h × (1 + r) of the expanded candidate area are, where the value of r ranges from 0.1 to 0.
3.
5. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S4, the semantic segmentation model is a DeepLabV3+ variant segmentation model, which includes a GA geometric attention module to enhance the model's pixel-level recognition ability for defect boundaries, slender structures, and irregular shapes.
6. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S5, the morphological features include the defect mask area, the number of connected components, the aspect ratio, the contour shape, the defect length, and the defect width. The comprehensive determination includes: when the defect mask area is greater than a preset area threshold, the maximum connected component satisfies a preset shape condition, and the segmentation confidence is greater than a preset confidence threshold, the current suspected defect candidate box is determined to be a real defect.
7. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S5, the final defect detection result includes: defect category, defect location, defect pixel-level segmentation mask, defect area, defect length, defect width, defect confidence level, and product OK / NG determination result.
8. The method for detecting tire surface defects with a low false negative rate according to claim 1, characterized in that: In step S6, the final defect detection result is output to a display terminal, host computer, PLC, MES or quality management system, thereby triggering an alarm, rejection, re-inspection or manual review process for the defective tire.
9. A low-miss-rate detection system for tire surface defects, characterized in that, include: The image acquisition and defect annotation module is used to acquire images of the tire surface and to perform target detection box annotation and pixel-level segmentation mask annotation on the defect areas in the images. The sample augmentation module is used to perform sample augmentation on the labeled training samples to generate augmented training samples with diverse forms. The model training and candidate box generation module is used to input the enhanced training samples into the target detection model for training. During the detection phase, the image of the tire surface to be detected is input into the target detection model, and a candidate box for suspected defects is output with a low confidence threshold lower than the preset standard threshold. The ROI expansion and segmentation confirmation module is used to expand the suspected defect candidate box, crop out the ROI that retains the surrounding context information of the defect, and input the ROI into the semantic segmentation model to output a defect pixel-level segmentation mask. The defect comprehensive judgment module is used to comprehensively judge the suspected defect candidate box based on the morphological features and segmentation confidence of the defect pixel-level segmentation mask, and output the final defect detection result. The result output and closed-loop management module is used to output the final defect detection results and perform quality statistics and closed-loop data management based on the final defect detection results.