Composite material infrared defect detection method and system based on multi-stream enhanced semi-supervised learning network

Abstract

The invention discloses a composite material infrared defect detection method and system based on a multi-stream enhanced semi-supervised learning network, which are applied to the technical field of nondestructive testing, and the method comprises the following steps: inputting composite material infrared defect label data to the multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model; the method comprises the following steps: carrying out step-by-step segmentation on infrared defects of a composite material by utilizing Pixel-Boosting through background pixel simulation based on sequential padding, carrying out overall visualization, and carrying out infrared defect positioning by utilizing an edge tracking algorithm. According to the method, the defects of an existing infrared thermal imaging nondestructive testing technology in the aspects of data set acquisition, semantic segmentation, defect identification and quantification are overcome, so that the infrared detection precision is effectively improved, and a new thought is provided for defect detection of the carbon fiber composite serving in the aerospace field.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, and in particular to a composite material infrared defect detection method and system based on a multi-stream enhanced semi-supervised learning network. Background Art

[0002] As one of the nondestructive testing technologies, infrared thermal imaging nondestructive testing technology is not limited by the test piece material compared to eddy current testing and other technologies. It can effectively detect ferromagnetic, conductive, non-ferromagnetic, and non-conductive materials.

[0003] Most algorithms used in existing infrared thermal imaging nondestructive testing technologies are used to segment and quantify a specific type of material. However, there are still significant differences when segmenting data from different types of specimens, which can easily lead to false detections and missed detections. Feature extraction algorithms can reduce the dimension of infrared videos and obtain thermal maps containing effective defect information, but the above images alone cannot determine which ROI areas represent defects, which requires the use of semantic segmentation algorithms to achieve quantitative segmentation of defects. Therefore, the further development of infrared thermal imaging nondestructive testing technology faces the challenge of insufficient robustness of segmentation and quantification algorithms.

[0004] In addition, the various challenges encountered during data collection have led to a serious shortage of data, which is also the main bottleneck in the development of current infrared non-destructive testing segmentation algorithms. Most of the current algorithms are based on the semantic segmentation framework of deep learning, and their demand for a large amount of experimental data is particularly urgent. Although a variety of algorithms have been used to try to alleviate the difficulties caused by data scarcity, the actual effect is still poor. At present, the way to alleviate the problem of insufficient data volume is to combine artificially simulated defects with natural defects, but due to cost factors, it is difficult to obtain artificial defects on a large scale. For carbon fiber composite materials in the aerospace field, due to the particularity of the industry, it is even more difficult to publicly obtain the detection data of natural defects formed during service.

[0005] In traditional semantic segmentation datasets, there is a clear difference between the segmented objects and the background. For example, in intelligent driving, the semantic information of vehicles and pedestrians is quite different. In the field of infrared thermal imaging nondestructive testing, infrared thermal imagers capture the internal heat dissipation process of materials, and the detected temperature difference information is subtle. In addition, due to the limitations of the detection accuracy of the thermal imager itself, infrared images are often accompanied by significant light pollution, which not only affects the image quality, but may also cover up those sensitive area defects with small but critical temperature differences, thus bringing challenges to the detection effect.

[0006] Therefore, how to provide a method to overcome the shortcomings of existing infrared thermal imaging nondestructive testing technology in data set acquisition, semantic segmentation, defect identification and quantification, so as to effectively improve the accuracy of infrared detection and provide new ideas for carbon fiber composite material defect detection serving the aerospace field. The composite material infrared defect detection method and system based on multi-stream enhanced semi-supervised learning network is a key problem that technical personnel in this field urgently need to solve. Summary of the invention

[0007] In view of this, the present invention proposes a composite material infrared defect detection method and system based on a multi-stream enhanced semi-supervised learning network.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A composite material infrared defect detection method based on a multi-stream enhanced semi-supervised learning network, comprising:

[0010] Step 1: Input composite material infrared defect label data to the multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model;

[0011] Step 2: Pixel-Boosting is used to perform step-by-step segmentation of composite material infrared defects through background pixel simulation based on sequential padding, and the overall visualization is performed, and the infrared defect positioning is performed using the edge tracking algorithm.

[0012] Optionally, in step 1, the method for obtaining the composite material infrared defect label data is specifically as follows:

[0013] The composite material is subjected to infrared thermal excitation irradiation, and an infrared imager is used to collect infrared images of the composite material after irradiation. A heat gradient is generated according to the law of thermal radiation and the unevenness of energy conduction, and infrared defects are marked;

[0014] The infrared image after infrared defect annotation is downsampled and PCA processed to obtain the final composite material infrared defect label data.

[0015] Optionally, in step 1, the teacher model and student model of the multi-stream enhanced semi-supervised learning network are both DeepLabv3+ models based on ResNet50. Specific improvements include:

[0016] The improvement of the enhancement branch uses multiple enhancement flows, including: weak enhancement flow and strong enhancement flow;

[0017] Apply two predefined image-level strong perturbations on the original image and insert a dropout layer on the weakly perturbed image features;

[0018] Separated different levels of enhancement into separate streams;

[0019] Dual independent strong views are randomly sampled from the augmented pool and fed into the student model in parallel, and supervised by the shared weak view.

[0020] Optionally, in step 1, the multi-stream enhanced semi-supervised learning network is trained, specifically:

[0021] The composite material infrared defect label data are divided into 50% each, and semi-supervised learning of feature perturbations and fully supervised learning of pseudo labels are performed respectively until the F-score of the multi-stream enhanced semi-supervised learning network meets the preset threshold, and Focal-loss is used as the loss function of the multi-stream enhanced semi-supervised learning network.

[0022] Optional, the F-score of the multi-stream enhanced semi-supervised learning network is as follows:

[0023]

[0024] Among them, β is a parameter; precision is the precision rate, which indicates the proportion of pixels classified as infrared defect pixels that are real infrared defect pixels; recall is the recall rate, which indicates the proportion of the number of actual real infrared defects in the predicted infrared defects to the positive samples in all samples.

[0025] Optionally, in step 2, pixel-boosting is used to perform step-by-step segmentation of composite material infrared defects by simulating background pixels based on sequential padding, specifically:

[0026] The segmented infrared defect positions are simulated with background pixels based on sequential padding to ensure that the previous layer only processes the unsegmented infrared defects when inputting the signal to the next layer.

[0027] Optional, background pixel simulation based on sequential padding, specifically:

[0028] Use the Min-Enclosing-Circle function in OpenCV to calculate the minimum circumscribed circle of the detected defect contour and determine the center and radius of the minimum circumscribed circle;

[0029] Draw a new circle with the minimum circumscribed circle as the center, using the radius doubled, and intercept the annular area between the new circle and the minimum circumscribed circle;

[0030] The pixel values ​​in the ring are recorded, and a mask is generated using the pixel values ​​in the ring. The pixel values ​​in the mask are used to fill the defective positions in sequence. When the number of pixels in the ring area is insufficient, these pixels are repeatedly used for filling.

[0031] Optionally, in step 2, the results of the step-by-step segmentation are visualized as a whole, specifically:

[0032] The infrared defects segmented step by step are uniformly spliced, and reconstruction transformation is performed to obtain an infrared reconstructed image reflecting the prominent characteristic areas of the infrared defects, thereby obtaining an overall visualized infrared defect.

[0033] Optionally, in step 2, an edge tracking algorithm is used to locate infrared defects, specifically:

[0034] Defect edge detection: Use edge detection algorithms to locate points in the image where the pixel value change is greater than a preset threshold, and obtain the edge points of the object;

[0035] Defect contour tracking: Take each object edge point as the starting point, move to another edge point along the edge direction, and connect the edge points. Repeat the edge point movement and edge point connection operations until returning to the starting point to obtain all defect closed contours;

[0036] Defect contour location: The algorithm is used to represent each defect closed contour as a list of points, which is used as a pixel list of the infrared defect contour in the predicted image to locate the infrared defect position in the original image.

[0037] The present invention also provides a composite material infrared defect detection system based on a multi-stream enhanced semi-supervised learning network using a composite material infrared defect detection method based on a multi-stream enhanced semi-supervised learning network, comprising:

[0038] Composite material infrared defect prediction model training module: used to input composite material infrared defect label data to a multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model;

[0039] Infrared defect positioning module: It is used to segment the infrared defects of composite materials step by step through background pixel simulation using Pixel-Boosting, and to visualize the whole, and to locate the infrared defects using edge tracking algorithm.

[0040] It can be seen from the above technical solutions that, compared with the prior art, the present invention proposes a composite material infrared defect detection method and system based on a multi-stream enhanced semi-supervised learning network. By using a composite material infrared defect prediction model based on a multi-stream enhanced semi-supervised learning network, the image-level and feature-level enhancements are combined to further expand the original enhancement space, improve the learning ability and overall performance of the model under complex enhancement conditions, and use F-score as the training completion index of the model, and use Focal-loss as the loss function of the multi-stream enhanced semi-supervised learning network by comparison and selection, which further ensures the accuracy of the model. Furthermore, the background pixel simulation based on sequential padding obtained by comparison and selection of Pixel-Boosting is used to perform step-by-step segmentation of composite material infrared defects, fully mine the multi-scale feature information of the image, and perform overall visualization, and use the edge tracking algorithm to locate infrared defects. By combining the results of the above algorithm with the positioning framework, the automatic quantification of composite material infrared non-destructive detection defects is realized, overcoming the shortcomings of existing infrared thermal imaging non-destructive detection technology in data set acquisition, semantic segmentation, defect recognition, and quantification, thereby effectively improving the infrared detection accuracy and providing new ideas for carbon fiber composite material defect detection serving the aerospace field. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0042] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0043] Figure 2 Schematic diagram of the multi-stream enhanced semi-supervised learning network framework of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Embodiment 1:

[0046] Embodiment 1 of the present invention discloses a composite material infrared defect detection method based on a multi-stream enhanced semi-supervised learning network, such as Figure 1 As shown, including:

[0047] Step 1: Input composite material infrared defect label data to the multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model;

[0048] The method of obtaining composite material infrared defect label data is as follows:

[0049] The composite material is subjected to infrared thermal excitation irradiation (excitation methods include: pulse, short-time, long-time, and continuous irradiation), and an infrared imager is used to collect infrared images of the composite material after irradiation. A heat gradient is generated according to the law of thermal radiation and the unevenness of energy conduction, and infrared defects are marked (Labelme is used to mark and generate corresponding PNG annotation files);

[0050] The infrared image after infrared defect annotation is downsampled and PCA processed to obtain the final composite material infrared defect label data.

[0051] The teacher model and student model of the multi-stream enhanced semi-supervised learning network are both DeepLabv3+ models based on ResNet50, such as Figure 2 As shown in the figure, the specific improvements include: improvements to the enhancement branch, using multiple enhancement streams, including weak enhancement streams and strong enhancement streams. The traditional FixMatch method only uses image-level enhancement, which limits the model's ability to explore a wider enhancement space. To solve this problem, a unified enhancement framework is designed that combines image-level and feature-level enhancements to further expand the original enhancement space. Specifically, two predefined image-level strong perturbations are applied to the original image, and a simple dropout layer is inserted on the extracted weakly perturbed image features. In order to avoid a single enhancement stream being too difficult to learn, enhancements of different levels are divided into independent streams. This design aims to enable each enhancement stream to learn in its own enhancement space, improving the overall performance of the model. In order to make better use of the enhancement space, a method of randomly sampling dual independent strong views from the enhancement pool is proposed, which is input into the student model in parallel and supervised by shared weak views. This method can further enhance the learning ability of the model in complex enhancement situations.

[0052] The multi-stream enhanced semi-supervised learning network is trained as follows:

[0053] The composite material infrared defect label data is divided into 50% each, and semi-supervised learning of feature perturbations and fully supervised learning of pseudo labels are performed respectively until the F-score of the multi-stream enhanced semi-supervised learning network meets the preset threshold (the preset threshold is set to F-score ≥ 90%), and Focal-loss is used as the loss function of the multi-stream enhanced semi-supervised learning network.

[0054] During the training process, the network will use a portion of labeled data (teacher model) and unlabeled data (student model) for joint training. In this way, the network can use the information of unlabeled data to improve the performance of the model based on limited labeled data. Usually, the teacher model generates pseudo labels, and the student model is trained based on these pseudo labels. At the same time, multiple enhancement flows are used to enrich the learning content of the model to improve its generalization ability under different conditions.

[0055] Data partitioning: Composite material infrared defect label data is divided into two parts, one for semi-supervised learning of feature perturbation, and the other for full-supervised learning of generating pseudo labels. The "50% each" here means that the data is divided into two parts, one for full-supervised learning (generating pseudo labels) and the other for semi-supervised learning (feature perturbation). This division is to ensure that there are both labeled and unlabeled data in the training process, so as to take advantage of semi-supervised learning.

[0056] Semi-supervised learning: In the semi-supervised learning process, unlabeled data is used to train the model after feature perturbation. The goal of this part of the training is to enable the model to learn useful information on unlabeled data. Feature perturbation can be done in many ways, such as image enhancement, in order to make the model more robust and able to handle a variety of changing data.

[0057] Fully supervised learning (generating pseudo labels): A portion of the labeled data is used for fully supervised learning to train a preliminary model. This preliminary model will generate pseudo labels, which will be used to guide model training in the semi-supervised learning process.

[0058] Consistency regularization: In the semi-supervised learning process, consistency regularization is used to ensure that the model output remains consistent under small perturbations of the input data. This helps improve the stability and accuracy of the model.

[0059] Training termination condition: The training process will continue until the F-score of the multi-stream enhanced semi-supervised learning network reaches the preset threshold. F-score is an indicator that comprehensively considers precision and recall, and is used to measure the overall performance of the model. When the F-score meets the preset threshold, it means that the performance of the model in detecting infrared defects of composite materials has reached the expected level.

[0060] Determination of loss function: Through comparative experiments, the effects of three loss functions, BCE, Focal-loss and OHEM, were compared. The results show that Focal-loss and OHEM perform better in dealing with the problem of class imbalance because they reduce the proportion of invalid negative samples in training, allowing the model to focus more on feature mining of positive samples. However, since OHEM needs to calculate the loss values ​​of all samples in each training step and select the top-k samples for backpropagation, this will increase the computational cost. Therefore, Focal-loss was finally selected as the loss function of the multi-stream enhanced semi-supervised learning network.

[0061] The F-score of the multi-stream enhanced semi-supervised learning network is as follows:

[0062]

[0063] Among them, β is a parameter; precision is the precision rate, which indicates the proportion of pixels classified as infrared defect pixels that are real infrared defect pixels; recall is the recall rate, which indicates the proportion of the number of actual real infrared defects in the predicted infrared defects to the positive samples in all samples.

[0064] Step 2: Use Pixel-Boosting to perform step-by-step segmentation of composite material infrared defects through background pixel simulation based on sequential padding, fully mine the multi-scale feature information of the image, and perform overall visualization, and use the edge tracking algorithm to locate infrared defects.

[0065] The Pixel-Boosting framework helps the model better identify and segment defects in infrared images by simulating background pixels. Specifically, it uses different background pixel simulation schemes, such as Gaussian filtering simulation, mean simulation, sequential padding simulation, and Gaussian padding simulation, to gradually decouple defect information and improve detection accuracy. After completing defect segmentation, the module also uses edge tracking algorithms to accurately locate defects. Edge tracking algorithms can help determine the exact boundaries of defects, which is crucial for further analysis of the nature and size of defects.

[0066] Gaussian filter simulation: Gaussian filtering is a common image processing method that uses a Gaussian kernel generated by a Gaussian function to convolve an image to blur the image. In the context of infrared nondestructive testing, Gaussian filter simulation essentially treats defective pixels as noise. This method smoothes the image by convolving the Gaussian kernel to make the defect edge less obvious, thereby attempting to simulate background pixels. However, since Gaussian filtering retains the direction of the original defect edge and the result after weighted averaging may differ greatly from the background pixels, the effect is limited.

[0067] Mean simulation: The mean simulation method is to calculate the average value of global pixels after locating the defect position. Since there is a significant imbalance between background pixels and defect pixels, theoretically, the average value is directly taken, and the impact of defect pixels is small. The defect area will be covered by a mask of the same size as the defect, and the pixel value of the mask is set to the average value of the background pixels and the defect pixels. This method performs well for small defects, but is not ideal for large defects because the pyramid module of the model requires multi-scale feature information, and mean simulation may cause the model to fail to correctly identify the location of the defect.

[0068] Sequential padding simulation: This method refers to the padding operation in convolutional neural networks, and its purpose is to keep the image size unchanged. Specifically, the edge tracking algorithm is first used to locate the defect contour and calculate the minimum circumscribed circle of the defect contour. Then a circle is drawn with the center of the circle and the radius doubled, the annular area between the two circles is intercepted, and the pixel values ​​within the annular area are recorded. Use these pixel values ​​to generate a mask and fill the pixels at the defect location in sequence. If there are not enough pixels, the filling is repeated. Sequential padding simulation performs better on large defects. Although slight demarcation lines may appear, the model is usually still able to treat these areas as background.

[0069] Gaussian padding simulation: The steps of Gaussian padding simulation are similar to sequential padding, but after intercepting the circular area, it uses a set of simulated pixel values ​​generated based on Gaussian distribution to fill the defect area. First, the mean and variance of the pixels in the circular area are calculated, and then the Gaussian distribution is simulated to generate a set of pixel values. The problem with this method is that the pixel distribution of infrared thermal images usually does not follow the Gaussian distribution, so in larger defects, Gaussian padding may cause the dark pixels to be too densely distributed.

[0070] After comparing the experimental results, we finally chose sequential padding simulation as the optimal solution because it can better simulate background pixels and avoid errors in the subsequent segmentation steps. In this way, the model can more accurately identify and segment defects in infrared images.

[0071] Pixel-Boosting is used to perform step-by-step segmentation of composite material infrared defects through background pixel simulation based on sequential padding, specifically:

[0072] The segmented infrared defect positions are simulated with background pixels based on sequential padding to ensure that the previous layer only processes the unsegmented defects when inputting the signal to the next layer. This allows subsequent input images to contain fewer defects, and the corresponding segmentation model needs to pay attention to fewer defects. This allows full use of the diverse results brought by different enhancement streams to achieve more accurate segmentation results.

[0073] Background pixel simulation based on sequential padding, specifically:

[0074] Use the Min-Enclosing-Circle function in OpenCV to calculate the minimum circumscribed circle of the detected defect contour and determine the center and radius of the minimum circumscribed circle;

[0075] Draw a new circle with the minimum circumscribed circle as the center, using the radius doubled, and intercept the annular area between the new circle and the minimum circumscribed circle;

[0076] The pixel values ​​in the ring are recorded, and a mask is generated using the pixel values ​​in the ring. The pixel values ​​in the mask are used to fill the defective positions in sequence. When the number of pixels in the ring area is insufficient, these pixels are repeatedly used for filling.

[0077] The results of the level-by-level segmentation are visualized as a whole, specifically:

[0078] The infrared defects segmented step by step are uniformly spliced, and reconstruction transformation is performed to obtain an infrared reconstructed image reflecting the prominent characteristic areas of the infrared defects, thereby obtaining an overall visualized infrared defect.

[0079] The edge tracking algorithm is used to locate defects, specifically:

[0080] Defect edge detection: Use edge detection algorithms (e.g. Canny edge detection algorithm) to locate points in the image where the pixel value change is greater than a preset threshold to obtain the edge points of the object;

[0081] Defect contour tracking: Take each object edge point as the starting point, move to another edge point along the edge direction, and connect the edge points. Repeat the edge point movement and edge point connection operations until returning to the starting point to obtain all defect closed contours;

[0082] Defect contour location: The algorithm is used to represent each defect closed contour as a list of points, which is used as a pixel list of the infrared defect contour in the predicted image to locate the infrared defect position in the original image.

[0083] As shown in Table 1, the F-score comparison results of the Pixel-Boosting proposed in the present invention and other existing algorithms on four different specimens are shown. As can be seen from Table 1, Pixel-Boosting achieved the highest F-score, with an average F-score of 90.96, and performed equally well in the four specimens. In the fully supervised model, FCN's performance in the field of defect detection is disastrous, with an average F-score of only 45.0. This may be because FCN uses transposed convolution, which causes the output feature map resolution to decrease. Therefore, FCN performs poorly when dealing with small targets or tasks that require fine segmentation. The average F-score of DeeplabV3+ reached 84.24, and it performed very well on different specimens. In the semi-supervised model, U2PL also has a relatively good performance, with an average F-score of 81.76. Among them, the best effect of all models was achieved on specimen 4, with an F-score of 96.71. However, since the framework separates reliable and unreliable pixels by predicting entropy, pushes each unreliable pixel to a category queue composed of negative samples, and uses unreliable candidate pixels to train the model, for specimens 1 and 3 with many defects and a lot of noise, unreliable pixels play a smaller role, so their F-scores are only 74.64 and 70.22. Mean-teacher has the worst effect in the semi-supervised model, which is also a predictable result, because Mean-teacher uses mean square error loss on unlabeled data. For unbalanced data sets, MSE may make the model more inclined to optimize the dominant category (background pixels), which is not suitable for the binary classification problem of defect detection. However, FixMatch has only two enhancement streams, strong and weak, so it performs poorly in the face of complex noise in infrared nondestructive testing. Its average F-score is 75.12, and even as low as 62.88 in the case of curved test piece 3.

[0084] Table 1 F-score comparison results of different algorithms on four different specimens

[0085]

[0086] In general, the algorithm proposed in the present invention performs well in all four specimens. In comparison with the fully supervised and semi-supervised networks, it outperforms the best performing DeeplabV3+ and U2PL by 6.72 and 9.2 respectively. Combined with the visualization results, it is proved that the Pixel-Boosting proposed in the present invention can effectively segment the defects of the infrared thermal images of composite materials and has good defect detection performance.

[0087] Embodiment 2:

[0088] Embodiment 2 of the present invention discloses a composite material infrared defect detection system based on a multi-stream enhanced semi-supervised learning network using a composite material infrared defect detection method based on a multi-stream enhanced semi-supervised learning network, comprising:

[0089] Composite material infrared defect prediction model training module: used to input composite material infrared defect label data to a multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model;

[0090] Infrared defect positioning module: It is used to segment the infrared defects of composite materials step by step through background pixel simulation using Pixel-Boosting, and to visualize the whole, and to locate the infrared defects using edge tracking algorithm.

[0091] The embodiment of the present invention discloses a composite material infrared defect detection method and system based on a multi-stream enhanced semi-supervised learning network. By using a composite material infrared defect prediction model based on a multi-stream enhanced semi-supervised learning network, image-level and feature-level enhancements are combined to further expand the original enhancement space, improve the learning ability and overall performance of the model under complex enhancement conditions, and use F-score as the training completion index of the model, and use Focal-loss as the loss function of the multi-stream enhanced semi-supervised learning network by comparison and selection, so as to further ensure the accuracy of the model. Furthermore, the background pixel simulation based on sequential padding obtained by comparison and selection of Pixel-Boosting is used to perform step-by-step segmentation of composite material infrared defects, fully mine the multi-scale feature information of the image, and perform overall visualization, and use the edge tracking algorithm to locate infrared defects. By combining the results of the above algorithm with the positioning framework, the automatic quantification of composite material infrared non-destructive detection defects is realized, and the shortcomings of existing infrared thermal imaging non-destructive detection technology in data set acquisition, semantic segmentation, defect recognition, and quantification are overcome, thereby effectively improving the infrared detection accuracy and providing new ideas for carbon fiber composite material defect detection serving the aerospace field.

[0092] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0093] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network, characterized in that: include: Step 1: Input composite material infrared defect label data to the multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model; Step 2: Pixel-Boosting is used to perform step-by-step segmentation of composite material infrared defects through background pixel simulation based on sequential padding, and the overall visualization is performed, and the infrared defect positioning is performed using the edge tracking algorithm.

2. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 1 is characterized in that: In step 1, the method for obtaining the composite material infrared defect label data is specifically as follows: The composite material is subjected to infrared thermal excitation irradiation, and an infrared imager is used to collect infrared images of the composite material after irradiation. A heat gradient is generated according to the law of thermal radiation and the unevenness of energy conduction, and infrared defects are marked; The infrared image after infrared defect annotation is downsampled and PCA processed to obtain the final composite material infrared defect label data.

3. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 1 is characterized in that: In step 1, the teacher model and student model of the multi-stream enhanced semi-supervised learning network are both DeepLabv3+ models based on ResNet50, and the specific improvements include: The improvement of the enhancement branch uses multiple enhancement flows, including: weak enhancement flow and strong enhancement flow; Apply two predefined image-level strong perturbations on the original image and insert a dropout layer on the weakly perturbed image features; Separated different levels of enhancement into separate streams; Dual independent strong views are randomly sampled from the augmented pool and fed into the student model in parallel, and supervised by the shared weak view.

4. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 1 is characterized in that: In step 1, the multi-stream enhanced semi-supervised learning network is trained, specifically: The composite material infrared defect label data is divided into 50% each, and semi-supervised learning of feature perturbation and fully supervised learning of pseudo labels are performed respectively until the F-score of the multi-stream enhanced semi-supervised learning network meets the preset threshold, and Focal-loss is used as the loss function of the multi-stream enhanced semi-supervised learning network.

5. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 4 is characterized in that: The F-score of the multi-stream enhanced semi-supervised learning network is as follows: Among them, β is a parameter; precision is the precision rate, which indicates the proportion of pixels classified as infrared defect pixels that are real infrared defect pixels; recall is the recall rate, which indicates the proportion of the number of actual real infrared defects in the predicted infrared defects to the positive samples in all samples.

6. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 1 is characterized in that: In step 2, pixel-boosting is used to perform step-by-step segmentation of composite material infrared defects by simulating the background pixels based on sequential padding, specifically: The segmented infrared defect positions are simulated with background pixels based on sequential padding to ensure that the previous layer only processes the unsegmented infrared defects when inputting the signal to the next layer.

7. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 6 is characterized in that: The background pixel simulation based on sequential padding is specifically as follows: Use the Min-Enclosing-Circle function in OpenCV to calculate the minimum circumscribed circle of the detected defect contour, and determine the center and radius of the minimum circumscribed circle; Taking the center of the minimum circumscribed circle as the center, draw a new circle using the radius doubled, and intercept the annular area between the new circle and the minimum circumscribed circle; The pixel values ​​in the ring are recorded, and a mask is generated using the pixel values ​​in the ring. The pixel values ​​in the mask are used to fill the defective positions in sequence. When the number of pixels in the ring area is insufficient, these pixels are repeatedly used for filling.

8. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 1 is characterized in that: In step 2, the results of the step-by-step segmentation are visualized as a whole, specifically: The infrared defects segmented step by step are uniformly spliced, and reconstruction transformation is performed to obtain an infrared reconstructed image reflecting the prominent characteristic areas of the infrared defects, thereby obtaining an overall visualized infrared defect.

9. The composite material infrared defect detection method based on multi-stream enhanced semi-supervised learning network according to claim 1 is characterized in that: In step 2, the edge tracking algorithm is used to locate infrared defects, specifically: Defect edge detection: Use edge detection algorithms to locate points in the image where the pixel value change is greater than a preset threshold, and obtain the edge points of the object; Defect contour tracking: Take each object edge point as the starting point, move to another edge point along the edge direction, and connect the edge points. Repeat the edge point movement and edge point connection operations until returning to the starting point to obtain all defect closed contours; Defect contour location: The algorithm is used to represent each defect closed contour as a list of points, which is used as a pixel list of the infrared defect contour in the predicted image to locate the infrared defect position in the original image.

10. A composite material infrared defect detection system based on a multi-stream enhanced semi-supervised learning network using the composite material infrared defect detection method based on a multi-stream enhanced semi-supervised learning network as described in any one of claims 1 to 9, characterized in that: include: Composite material infrared defect prediction model training module: used to input composite material infrared defect label data to a multi-stream enhanced semi-supervised learning network for training to obtain a composite material infrared defect prediction model; Infrared defect positioning module: It is used to segment the infrared defects of composite materials step by step through background pixel simulation using Pixel-Boosting, and to visualize the whole, and to locate the infrared defects using edge tracking algorithm.

Images