Tire X-ray image bubble defect detection method based on noise label correction

Abstract

The invention discloses a tire X-ray image bubble defect detection method based on noise label correction. The method comprises a noise label correction stage and a defect detection stage. In the correction stage, firstly, a noise-containing training set is divided into two subsets, two detection models are obtained through iterative training, and cross prediction is carried out; secondly, adopting momentum updating to gradually correct the coordinates of the labeling box, wherein a weight parameter omega is gradually increased along with the number of iterations; and finally, outputting the corrected data set and retraining to obtain an optimization model. In the detection stage, the optimization model is used for carrying out bubble defect detection on the new X-ray image, and the defect position and the confidence coefficient are output. According to the method, the marking noise can be effectively corrected, the accuracy and robustness in tire bubble defect detection are remarkably improved, and excellent performance can still be kept even under the condition that the noise reaches up to 60%.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and industrial inspection technology, specifically to a method, system, and equipment for detecting air bubble defects in tire X-ray images based on noise label correction. Background Technology

[0002] Tire X-ray images show good representation of some defect types, such as steel belt structure and tire cord skeleton materials. However, they are not very effective at identifying tire bubbles, for example, the difference in grayscale values ​​is not obvious compared to normal areas. Currently, tire X-ray image-based defect identification methods primarily rely on manual labeling. Two common errors occur during manual labeling: incorrect category labeling and inaccurate bounding boxes. Both types of label noise can lead to a decrease in model accuracy after training. Of these two errors, inaccurate bounding boxes are more likely to occur, especially for tire bubbles, whose images are already quite similar in grayscale value to normal areas. When a large number of labeled samples with inaccurate bounding boxes appear, misidentification and misjudgment of tire bubble defects become highly likely. Therefore, the accuracy requirements for labeling tire bubbles are high. In fact, tire manufacturing employees often experience a decrease in labeling accuracy for these difficult-to-identify features. Therefore, finding a more effective method to correct such noise labels has become an important part of the method for identifying bubble-like defects using tire X-ray images. Summary of the Invention

[0003] This invention provides a method for detecting air bubble defects in tire X-ray images based on noise label correction. The core of this method lies in gradually correcting noise labels in the training data through an iterative cross-training and momentum update mechanism. The specific scheme is as follows:

[0004] First, the training dataset containing noisy labels is randomly divided into two subsets with equal sample sizes. Then, detection models are trained on these two subsets respectively, and the training process and performance are monitored using a noise-free validation set. Next, the trained models are used to perform cross-prediction on the other subset. Afterward, momentum updates are performed based on the difference between the prediction results and the current labels, gradually adjusting the label information. Finally, the algorithm stops iterating based on a designed termination condition. If the termination condition is not met, the next iteration continues; if the termination condition is met, the algorithm terminates, outputting the final calibrated dataset and the optimized detection model.

[0005] The specific implementation steps are as follows:

[0006] (1) Dataset definition

[0007] Assume a training dataset containing noise. It consists of input and labels, and is represented as Input here It can be various types of data, such as images, text features, etc., while labels... Since it's a single-class problem, it only contains information such as coordinates and dimensions. Furthermore, it has a noise-free validation set. It consists of input and label, and is represented as ,same It contains only information such as coordinates and dimensions, and is used to objectively evaluate model performance.

[0008] (2) Dataset partitioning

[0009] This article will use a training dataset containing noise. Randomly divide into two subsets of the same size. and The superscript indicates the iteration number. The dataset for the i-th iteration is represented as shown in Equation (2-1). This partitioning method helps to avoid the influence of model overfitting during data correction and enhances the robustness of the algorithm.

[0010] (2-1)

[0011] (3) Model training process

[0012] The model is trained and tested using a partitioned subset and a clean, noise-free validation set. Specifically, it utilizes... The model is obtained through training ,use The model is obtained through training ,in Calculate the model's average accuracy metric and use strategies such as early stopping to determine the optimal time to terminate model training, ensuring that the model has good generalization ability.

[0013] (4) Prediction process

[0014] The model is obtained through training. and Then, use these models to perform cross-prediction on another subset. Using the models... Predict a subset of A, i.e. Using models Predict subset B, i.e. This cross-prediction mechanism ensures that the predictions used for correction come from independently trained models, reducing bias.

[0015] The prediction results here and These are the model's predicted outputs for two subsets of inputs, which will be used in the subsequent momentum update process. The accuracy of the predictions depends on the model's training performance, which in turn is affected by the training data and training process mentioned earlier.

[0016] (5) Momentum update

[0017] Momentum update is one of the key steps in the algorithm presented in this paper, and its formula is as follows:

[0018] (2-2)

[0019] Here, ω is a parameter between 0 and 1, used to control the weight of the original label information and the prediction result in the updated label. In this way, the label information can be gradually adjusted to make it closer to the true label value.

[0020] Momentum updates give the label information updates a certain inertia. In each iteration, the updated label... and It not only takes into account the current forecast results and It also retains some of the tag information from the previous iteration. and This avoids drastic fluctuations in tag information during the update process, making the update process smoother.

[0021] The role of momentum updates is mainly reflected in two aspects. Firstly, it can use the model's predictions to correct noisy original label information. If the model's predictions are relatively accurate, momentum updates can gradually bring the label information closer to the true values. Secondly, momentum updates can improve the model's stability and generalization ability. By gradually adjusting the label information, the model can better adapt to different data conditions and reduce the risk of overfitting.

[0022] Updated tags , The dataset used to construct the (i+1)th iteration:

[0023] (2-3)

[0024] Then , Return to step 3 and repeat the model training process until the termination condition is met.

[0025] (6) Design of termination conditions

[0026] To ensure the algorithm stops iterating at an appropriate time, a termination condition needs to be designed. When... , The iteration should stop when the difference is small enough. The design of the algorithm's termination condition is crucial to both efficiency and performance. If the number of iterations is too small, the algorithm may struggle to converge the parameters; conversely, excessive iterations not only increase computational costs but may also lead to overfitting. By designing a reasonable termination condition, the algorithm's efficiency can be improved while maintaining model performance.

[0027] Compared with the prior art, the present invention has the following significant advantages:

[0028] This invention effectively corrects noisy labels in the training set through iterative cross-training and momentum updates. Even with an extreme noise ratio of up to 60%, it can still significantly improve the performance of the detection model and demonstrate strong robustness.

[0029] The detection model trained on the dataset corrected by the method of this invention has significantly improved key indicators such as average precision and recall on the independent test set, achieving accurate and comprehensive bubble defect detection.

[0030] The method of this invention solves the pain point of high cost in obtaining high-quality labeling data in the tire production environment. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the original annotation box;

[0032] Figure 2 and Figure 3 A comparison diagram illustrating the generation of annotation box noise;

[0033] Figure 4 Precision-recall curve for noisy data;

[0034] Figures 5-7 A graph showing the changing trends of various evaluation indicators during noise training;

[0035] Figure 8 This is an example of the noise label correction process. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The following embodiments are for illustrative purposes only and should not be considered as limiting the scope of the invention.

[0037] Noise label correction phase:

[0038] Step S1: Data Preparation and Preprocessing

[0039] The Pascal VOC format object detection dataset for tire bubble defects was converted to YOLO format. The converted dataset includes a training set of 10,000 images and a test set of 1,500 images. To ensure the effectiveness and accuracy of the model's test metrics, the test set annotations underwent multiple manual checks, verifying the position and category accuracy of each image to achieve near-perfect annotations, making it an ideal noise-free dataset. The test set remained unchanged and was used for validating and testing model performance metrics, while the original training set was primarily used to calculate the correction effect for noisy labels.

[0040] Step S2: Noise Label Simulation and Dataset Partitioning

[0041] YOLOv10 was chosen as the baseline model for the entire experiment. The original training set was used for training first, and the training results were recorded. The main parameters were set as follows: batch size of 128 (128*n for multi-GPU training, where n is the number of GPUs), number of epochs of 200, and input size of 640×640. The parameter settings for subsequent training were kept the same.

[0042] The method for generating a noisy dataset from the training set by artificially adding contamination labels is as follows:

[0043] a. The training set is randomly divided into three parts: S1 (30%), S2 (30%), and S3 (40%).

[0044] b. Noise concentration 30%: Pollution S1, S2 and S3 remain clean;

[0045] c. Noise concentration 60%: Pollution in S1 and S2, S3 remains clean;

[0046] The pollution algorithm is as follows:

[0047] a. Let the center coordinates of the original annotation box be x, y, and the width and height be w, h, as follows: Figure 1 As shown;

[0048] b. Set the new center point of the noise label box to Figure 1 A random point within the points, with its width and height randomly scaled to w. n h n A scaling factor is randomly selected between 0.5 and 2. An example of generating noise in the annotation box is shown below. Figure 2 and Figure 3 As shown;

[0049] (2-4)

[0050] Step S3: Iterative Training and Correction

[0051] Taking a noise set of 30% as an example, the same applies to a noise set of 60%.

[0052] 1. Model Training: A 30% noise set is used as the training set, and the test set is used as the validation set, serving as a benchmark. Early stopping is employed during training; training is stopped when the model's performance on the validation set no longer improves, to prevent overfitting. All training runs maintain strictly the same parameter settings: batch size of 128 (128*n for multi-GPU training, where n is the number of GPUs), 200 epochs, and input size of 640×640.

[0053] First, randomly divide the 30% noise set into two parts: noise set A0 and noise set B0. Then train them separately, using noise set A0 as the training set and the test set as the validation set to obtain model A0, and using noise set B0 as the training set and the test set as the validation set to obtain model B0. The same applies to the 60% noise set.

[0054] 2. Alternating evaluation and correction: Use model A0 to evaluate the 30% noise set B0 of the dataset and apply correction, and use model B0 to evaluate the 30% noise set A0 of the dataset and apply correction.

[0055] Correction method: For each label bounding box, find the best matching prediction box (the matching rule is the maximum intersection-union ratio; if there are no intersecting prediction boxes, ignore the correction step).

[0056] 3. Iterative correction process:

[0057] This embodiment performs three iterations of correction, with the weight parameter ω increasing with the number of iterations.

[0058] First iteration correction: Assuming an initial weight ω = 0.5, a weighted average correction is applied to the bounding box coordinates (x, y, w, h) to obtain the new training set noise set 30%A1 and noise set 30%B1.

[0059] (2-5)

[0060] Second iteration correction: Repeat training and correction, adjusting weight ω=0.6.

[0061] (2-6)

[0062] We obtain a new training set with 30% noise (A2) and 30% noise (B2).

[0063] Third iteration correction: Repeat the training and correction steps, adjusting the weights ω=0.7.

[0064] (2-7)

[0065] The iterative correction process terminates after all three preset iterations are completed. The final corrected training sets A3 and B3 are then merged and used as the corrected training set for subsequent model retraining.

[0066] Step S4: Model Retraining

[0067] The bubble defect detection model was retrained using the final corrected complete training dataset to obtain the optimized final detection model. The retraining process maintained the same parameter settings: batch size of 128 (128*n for multi-GPU training, where n is the number of GPUs), number of epochs of 200, and input size of 640×640, to ensure consistency of training conditions.

[0068] II. Defect Detection Stage

[0069] Step 5: Use the optimized bubble defect detection model obtained in Step S4 to detect bubble defects in the new tire X-ray image, and output the detection results including the bounding box coordinates of the bubble defect location and the detection confidence score. The detection results are visualized in the form of bounding boxes, including the bubble's center coordinates (x, y), width w, height h, and presence confidence score.

[0070] Performance verification and analysis:

[0071] To verify the effectiveness of the proposed method, five reference indicators were designed for evaluation, as follows:

[0072] a. Recall: Recall refers to the proportion of all positive samples that the model correctly identifies as positive.

[0073] b. Average Precision 50: The average precision at an Intersection over Union (IoU) threshold of 0.5. This metric measures the model's performance in the detection task.

[0074] c. Average accuracy 50-95: Average accuracy across cross-union ratios (CURRs) from 0.5 to 0.95. This metric evaluates the overall performance of the model across various CURR thresholds.

[0075] d. Distribution focus loss refers to the measurement of the accuracy of the bounding boxes and related information predicted by the model.

[0076] e. Precision: refers to the proportion of actual positive samples out of all predictions that the model classifies as positive.

[0077] f: Cross-Union Ratio (IoU) of prediction results for all samples: the average IoU between the model's predicted bounding box and the labeled bounding boxes in the current dataset.

[0078] g: Crossover Union Ratio of Contaminated Sample Prediction Results: The average IoU between the model's predicted bounding box and the current labeled bounding box on the sample portion initially injected with noise.

[0079] Using the test set as the validation set, we evaluated the convergence comparison of datasets with different levels of contamination (30% noise set and 60% noise set) in the YOLOv10-N model. Figure 4 Table 1 shows the performance on the noisy dataset. Without contaminated labels, the detection model achieved an average precision (AP) of 0.617 on the original dataset, indicating good detection performance. After adding 30% noisy data, the AP dropped to 0.566, the cross-union ratio (CUI) for all samples was 0.781, and the CUI for contaminated samples was only 0.269, indicating that noise significantly impacted the detection model's performance. After adding 60% noisy data, the AP dropped to 0.560, the CUI for all samples was 0.559, and the CUI for contaminated samples was only 0.265, indicating that noise further amplified the impact on the detection model's performance.

[0080] After calibration using the autopilot sampling strategy, the average precision of the 30% noise set (50%) improved to 0.677, the cross-union ratio (CUN) of the overall sample prediction results was 0.680, and the CUN of the contaminated sample prediction results improved to 0.582, significantly improving the performance of the detection model. The average precision of the 60% noise set (50%) further improved to 0.694, the CUN of the overall sample prediction results was 0.618, and the CUN of the contaminated sample prediction results improved to 0.492, significantly improving the robustness of the detection model.

[0081] It is worth noting that the cross-union ratio (CUI) of the overall sample prediction results in the corrected dataset decreased, which precisely proves the effectiveness of label correction. This is because the initial high CUI of 0.781 represents the matching degree between the model prediction and the inaccurate noisy labels; after correction, the labels become closer to the true values, and although the matching degree between the model prediction and the more accurate new labels (0.680) decreases in value, it represents a more realistic positioning accuracy.

[0082] Table 1 Performance of Noisy Datasets

[0083]

[0084] Figures 5-7 The various metrics of the training process are displayed. The evaluation results clearly show that as the number of training steps increases, the recall, mean precision (50%), mean precision (50-95%), and precision of each metric gradually increase and tend to stabilize. This indicates that the model gradually learns effective features during training, and its performance is improved, thus verifying the effectiveness of the method proposed in this chapter.

[0085] Table 2 shows a comparison of the datasets under different correction weights. Under different correction weights, the average accuracy and cross-union ratio (CUN) of each indicator on corrections 30 and 60 gradually increase with the increase of correction weight. When the correction weight increases to 0.7, the average accuracy of correction 30 is 0.677 and the CUN of contaminated sample predictions is 0.582, while the average accuracy of correction 60 is 0.694 and the CUN of contaminated sample predictions is 0.492, further validating the effectiveness of the method proposed in this chapter.

[0086] Table 2 Comparison of datasets under different correction weights

[0087]

[0088] Figure 8 This demonstrates an example of the noise label correction process. On a dataset with 30% noise correction, noise generated simultaneously in width, height, and center point was gradually corrected to obtain more accurate bounding boxes for width, height, and center point. Similarly, on a dataset with 60% noise correction, noise generated only in width and height was also gradually corrected to obtain accurate bounding boxes.

[0089] The effectiveness of the algorithm in noisy labeling scenarios was verified by artificially constructing two noisy labeled datasets: one with 30% noise and the other with 60% noise. The experiment with the 60% noise set is particularly significant—it simulates the algorithm's performance under extreme conditions by setting a mislabeling ratio as high as 60%. Traditionally, it is believed that when the mislabeling ratio in a dataset is too high, the model struggles to extract useful information from the noise and often fails to converge to a reasonable result. However, this experiment demonstrates that even under the extreme noise condition of 60%, the proposed algorithm can still significantly correct mislabels and recover discriminative feature representations.

[0090] More notably, the experimental results indirectly revealed potential labeling quality issues in the original dataset. Comparisons showed that the corrected models trained with 30% and 60% noise correction significantly outperformed the baseline model trained directly on the original dataset on independent test sets. This strongly supports the research hypothesis that the original dataset inevitably contains significant noise labels. It is precisely because of the algorithm's robust noise correction capability that the model trained on a subset with artificially injected controllable noise performs better than the model trained on the seemingly correct original dataset. This result precisely verifies the core value of the algorithm design: in real-world scenarios, even with large-scale noise pollution, it can effectively uncover the inherent regularities in the data, thereby improving the generalization ability of the detection model.

Claims

1. A method for detecting bubble defects in a tire X-ray image based on noise label correction, characterized by, The method comprises the following steps: (I) noise label correction stage: Step S1: acquire a tire X-ray image training data set, wherein the tire X-ray image contains a bubble defect area which is difficult to label due to an unobvious gray scale difference; construct a training data set containing noise labels ; wherein The tire X-ray image data can be The bounding box label for labeling the bubble defect position has noise in the inaccurate labeling box position; and a noise-free verification set is defined ;​ Step S2: splitting the training dataset into two subsets of equal number of samples randomly split into two subsets of equal number of samples and , where the upper index denotes the iteration number, the dataset of the i-th iteration; Step S3: iterative training and correction In the i-th iteration, using the subset Training to obtain a bubble defect detection model , using the subset Training to obtain a bubble defect detection model ; Using the model On the subset Making a prediction, i.e. ; using the model On the subset Making a prediction, i.e. ; Step S4: correct the label using the momentum update formula: (2-2) ω is a parameter between 0 and 1; updated label , Dataset for building the i+1 step iteration: (2-3) Then , The model training process is repeated back to step S3 until the termination condition is met; Output the final corrected training data set and the optimized bubble defect detection model; (II) defect detection stage: step S5: use the optimized bubble defect detection model to detect bubble defects in a new tire X-ray image, and output the detection result containing the bubble defect position bounding box coordinates and detection confidence.

2. The method of claim 1, wherein, The weight parameter ω increases with the increase of the number of iterations.

3. The method of claim 2, wherein, The initial value of the weight parameter ω is between 0.4 and 0.6, and gradually increases to 0.7 in subsequent iterations.

4. The method of claim 1, wherein, In the prediction process, for each original annotation box, the intersection-over-union is used as the matching criterion to find the maximum intersection-over-union prediction box for matching; if there is no intersecting prediction box, skip the correction step of the annotation box.

5. The method of claim 1, wherein, The bubble defect detection model is based on the YOLOv10 architecture.

6. The method of claim 1, wherein, The optimized bubble defect detection model is obtained by merging the final iteration corrected training sub-set and retraining using the merged complete corrected training data set.

7. The method of claim 1, wherein, The iteration termination condition is that a difference between the updated label , and the corresponding prediction result , is less than a preset threshold.

8. The method of claim 1, wherein, The training data set containing noise labels is constructed by artificially polluting part of the annotation boxes to generate noise from the original training set; the pollution includes randomly shifting the center coordinates of the annotation boxes and / or randomly scaling the width and height of the annotation boxes.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the electronic device to perform the method of any one of claims 1-8. The processor executes the program to implement the steps of the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method of any one of claims 1-8.

Images