Industrial CT hierarchical detection method, device, system and storage medium

Abstract

The present invention discloses an industrial CT hierarchical detection method, device, system, and storage medium, comprising obtaining a first two-dimensional image provided by a first industrial CT scanning device; detecting the first two-dimensional image based on a first detection model; when the first detection result is characterized as a pending result, obtaining a second two-dimensional image provided by a second industrial CT scanning device; detecting the second two-dimensional image based on a pre-trained second detection model; when the second detection result is characterized as a pending result, obtaining a three-dimensional image model provided by a third industrial CT scanning device; displaying the three-dimensional image model for manual analysis to obtain a third detection result; annotating the first two-dimensional image and the second two-dimensional image according to the third detection result, and adding the first two-dimensional image and the second two-dimensional image to the training data sets of the first detection model and the second detection model, respectively. The present invention integrates the advantages of different CT detection methods through a hierarchical detection system to achieve efficient and accurate product detection.

Description

Technical Field

[0001] The present invention relates to the field of automated detection technology, and in particular to an industrial CT hierarchical detection method, device, system and storage medium. Background Art

[0002] Industrial CT testing is crucial for quality control in industrial production, enabling non-destructive testing of product internal structures and timely detection of defects. However, current industrial CT testing faces numerous challenges, including limited application scenarios for different testing methods and difficulty balancing efficiency and accuracy.

[0003] Traditional inspection methods are ineffective for detecting minor defects and complex structures. Existing deep learning inspection networks are limited by the limited amount of data available in laboratories, and their accuracy and reliability need to be improved. During the inspection process, the inability to accurately distinguish between different defect levels and types can lead to over-inspection or missed inspections, increasing production costs and reducing efficiency. Summary of the Invention

[0004] The present invention aims to address at least one of the technical problems existing in the prior art. To this end, the present invention proposes a hierarchical industrial CT inspection method, device, system, and storage medium. This hierarchical inspection system integrates the advantages of different CT inspection methods to achieve efficient and accurate product inspection.

[0005] In a first aspect, an embodiment of the present invention provides an industrial CT hierarchical detection method, comprising:

[0006] Acquire a first two-dimensional image of a target product provided by a first industrial CT scanning device;

[0007] Detecting the first two-dimensional image based on a pre-trained first detection model to obtain a first detection result;

[0008] When the first detection result is characterized as pending, obtaining a second two-dimensional image of the target product provided by a second industrial CT scanning device, the second industrial CT scanning device having a higher theoretical spatial resolution than the first industrial CT scanning device;

[0009] Detecting the second two-dimensional image based on a pre-trained second detection model to obtain a second detection result;

[0010] When the second detection result indicates that the result is pending, obtaining a three-dimensional image model of the target product including internal structural features provided by a third industrial CT scanning device;

[0011] displaying the three-dimensional image model for manual analysis to obtain a third detection result;

[0012] The first two-dimensional image and the second two-dimensional image are annotated according to the third detection result, and the first two-dimensional image is added to the training data set of the first detection model and the second two-dimensional image is added to the training data set of the second detection model.

[0013] According to some embodiments of the present invention, the first detection model includes a first detection module and a first classification module, the first detection module is constructed based on the YOLOv11 model, the size of the anchor box of the YOLOv11 model is configured to be multiple of 16×16, 32×32, 64×64 and 128×128 pixels, and the first classification module is constructed based on the EfficientNet-B model, and the number of input channels of the EfficientNet-B model is configured to be a single channel.

[0014] According to some embodiments of the present invention, the first detection result includes a first defect location and a first confidence level, and the first two-dimensional image is detected based on the pre-trained first detection model to obtain the first detection result, and then the following steps are further included:

[0015] When the first defect position is not empty or the first confidence level is lower than a first preset threshold, the first detection result is characterized as pending.

[0016] According to some embodiments of the present invention, the second detection model includes a second detection module and a second classification module. The second detection module is constructed based on an enhanced YOLOv11 model. The enhanced YOLOv11 model uses a feature pyramid network to fuse shallow features and deep features. The shallow features are used to represent the feature layer with an image resolution of 1024×1024, and the deep features are used to represent the feature layer with an image resolution of 128×128 pixels. The second classification module is based on the EfficientNet-lite model.

[0017] According to some embodiments of the present invention, an SE module is added to the MBConv module of the EfficientNet-lite model.

[0018] According to some embodiments of the present invention, the second detection model is configured with an enhanced loss function during the training phase, and the calculation formula of the enhanced loss function is:

[0019] ,

[0020] Where L represents the detection loss of the second two-dimensional image, α δ Characterized as the defect scale weighting factor, L small Characterized as positioning loss of small size defects, where the small size defect is used to characterize the defect area less than 100μm2 defects.

[0021] According to some embodiments of the present invention, the calculation formula for the positioning loss of the small-size defect is:

[0022] ,

[0023] In the formula, IOU is used to represent the intersection-over-union ratio between the predicted box A and the real box B, and C is used to represent the minimum enclosing rectangle of the predicted box A and the real box B.

[0024] In a second aspect, an embodiment of the present invention provides an industrial CT hierarchical detection device, comprising:

[0025] A first acquisition module is used to acquire a first two-dimensional image of a target product provided by a first industrial CT scanning device;

[0026] A first determination module is configured to detect the first two-dimensional image based on a pre-trained first detection model to obtain a first detection result;

[0027] a second acquisition module, configured to, when the first detection result is characterized as pending, acquire a second two-dimensional image of the target product provided by a second industrial CT scanning device, the second industrial CT scanning device having a higher theoretical spatial resolution than the first industrial CT scanning device;

[0028] a second determining module, configured to detect the second two-dimensional image based on a pre-trained second detection model to obtain a second detection result;

[0029] a third acquisition module, configured to acquire, when the second detection result is characterized as pending, a three-dimensional image model of the target product including internal structural features provided by a third industrial CT scanning device;

[0030] a third determination module, configured to display the three-dimensional image model for manual analysis to obtain a third detection result;

[0031] A feedback module is used to annotate the first two-dimensional image and the second two-dimensional image according to the third detection result, and to add the first two-dimensional image to the training data set of the first detection model and to add the second two-dimensional image to the training data set of the second detection model.

[0032] In a third aspect, an embodiment of the present invention provides an industrial CT grading detection system, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is used to implement the above-mentioned industrial CT grading detection method when running the computer program.

[0033] In a fourth aspect, an embodiment of the present invention provides a storage medium, wherein a computer program is stored in the storage medium, and when the computer program is executed, the above-mentioned industrial CT hierarchical detection method is implemented.

[0034] The embodiments of the present invention have at least the following beneficial effects:

[0035] The first industrial CT scanning device is combined with the first detection model, the second industrial CT scanning device is combined with the second detection model, and the third industrial CT scanning device is combined with manual analysis to achieve hierarchical detection of target products. The first industrial CT scanning device and the first detection model are used to perform primary detection to quickly detect defective products and improve overall detection efficiency. For target products whose first detection results are characterized as pending, secondary or tertiary detection is performed as appropriate, integrating the advantages of different CT detection methods to achieve efficient and accurate detection of products. The two-dimensional image is annotated according to the detection results and added to the training data set of the detection model to optimize the detection model.

[0036] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0038] Figure 1 This is a flowchart of the steps of the industrial CT hierarchical detection method according to an embodiment of the present invention;

[0039] Figure 2 is a principle block diagram of a first detection model according to an embodiment of the present invention;

[0040] Figure 3 is a principle block diagram of a second detection model according to an embodiment of the present invention;

[0041] Figure 4 This is a principle block diagram of an industrial CT hierarchical detection device according to an embodiment of the present invention;

[0042] Figure 5 This is a principle block diagram of the industrial CT hierarchical detection system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0044] In the description of the present invention, "several" means one or more, "multiple" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, and "above," "below," and "within" are understood to include the number itself. The use of terms such as "first" and "second" is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0045] Industrial CT (Industrial Computerized Tomography) refers to computed tomography (CT) technology used in industry. Its fundamental principle is based on the attenuation and absorption characteristics of radiation in the object being inspected. Traditional inspection methods are ineffective for detecting small defects and complex structures. They typically utilize a single device or inspection network. When a potential defect is detected in a target product, manual re-inspection is often performed. Existing deep learning inspection networks are limited by the limited amount of data available within the laboratory, leaving room for improvement in detection accuracy and reliability. Manual re-inspection is labor-intensive, increasing production costs and reducing efficiency.

[0046] Please refer to Figure 1 This embodiment discloses an industrial CT hierarchical detection method, including steps S100 to S700. It should be noted that the numbering of the steps in this embodiment is only for the convenience of review and understanding, and does not limit the order in which the steps are executed. The content of each step is detailed below:

[0047] S100, obtaining a first two-dimensional image of a target product provided by a first industrial CT scanning device;

[0048] For example, the first industrial CT scanning device can employ high-precision industrial CT scanning equipment to perform a comprehensive scan of the target product. During the scanning process, scanning parameters, such as the X-ray tube voltage, tube current, and scan time, are appropriately adjusted based on the target product's material and shape, as well as the type of defects expected to be detected, to obtain a clear and complete X-ray projection image, where the X-ray projection image is a two-dimensional image. After obtaining the first two-dimensional image, data preprocessing can also be performed on the first two-dimensional image, such as removing noise interference and correcting the image's grayscale values, to improve image quality.

[0049] S200: Detect the first two-dimensional image based on a pre-trained first detection model to obtain a first detection result;

[0050] For example, compared to manual inspection, inspecting the first two-dimensional image based on a pre-trained first inspection model can quickly, accurately, and stably distinguish between good products and products with unclear test results in a batch of products, thereby improving inspection efficiency. Considering that the number of good products in industrial production typically far exceeds that of defective products, during pre-training of the first inspection model, industrial CT projection images containing features such as different materials (e.g., metal or non-metal) and different defect types (e.g., cracks or poor solder joints) can be collected as training samples to improve the detection accuracy of the first inspection model.

[0051] S300: When the first detection result indicates that the result is pending, obtain a second two-dimensional image of the target product provided by a second industrial CT scanning device, where the theoretical spatial resolution of the second industrial CT scanning device is higher than that of the first industrial CT scanning device;

[0052] For example, the second industrial CT scanner has a higher theoretical spatial resolution than the first industrial CT scanner, enabling more detailed scans of the target product, thereby obtaining higher-resolution two-dimensional images. In some application examples, the second industrial CT scanner can be a planar CT scanner, also known as a fan-beam CT scanner. Planar CT scanners, also known as fan-beam CT scanners, use a single row of detectors to receive fan-beam radiation, have high spatial resolution (single pixel size ≤ 5μm), and are capable of excellent two-dimensional tomographic imaging. Furthermore, the X-ray tube of a planar CT scanner can be placed in close proximity to the target product, enabling detailed scans of the target product. During the scanning process, parameters such as the scanning slice thickness and spacing are adjusted based on the product's size, shape, and inspection requirements to obtain high-quality planar CT scan image data.

[0053] S400: Detect the second two-dimensional image based on the pre-trained second detection model to obtain a second detection result;

[0054] Exemplarily, the first and second inspection models are two different inspection models that can operate independently. The second inspection model's input data comes from a second industrial CT scanner with a higher theoretical spatial resolution. This allows for clearer identification of defect characteristics in target products, improving inspection accuracy and reducing the number of products requiring manual re-inspection. If the second inspection result matches the first, inspection of the corresponding target product is terminated; otherwise, a more detailed scan is performed.

[0055] S500: When the second detection result indicates that the result is pending, obtain a three-dimensional image model of the target product including internal structural features provided by a third industrial CT scanning device;

[0056] For example, for a target product whose first and second inspection results are inconsistent, a third industrial CT scanning device is used to perform a three-dimensional scan of the target product to generate a three-dimensional image model containing the internal structural features of the target product. In some application examples, the third industrial CT scanning device uses a cone-beam CT device. The cone-beam CT device uses a cone-shaped beam of rays in conjunction with a planar array detector to capture a three-dimensional image model of the target product in a single rotation, and its scanning speed is higher than that of a planar CT scanning device. During the scanning process, the scanning parameters of the cone-beam CT device are adjusted according to the characteristics of the product and the possible types of defects. For example, for large products, the tube voltage and tube current of the X-ray tube of the cone-beam CT device are appropriately increased to increase the imaging range; for the detection of tiny defects, the scanning layer thickness and spacing are appropriately reduced to improve the image resolution.

[0057] S600, displaying the three-dimensional image model for manual analysis to obtain a third detection result;

[0058] For example, the three-dimensional image model collected by the third industrial CT scanning equipment can intuitively display the internal structural features of the target product, facilitating manual observation and analysis, such as observing the internal structural features of the target product, determining whether the target product has defects, and identifying various potential defects, and recording information such as the type, size, location and severity of the defects as the final detection result.

[0059] S700: Label the first two-dimensional image and the second two-dimensional image according to the third detection result, and add the first two-dimensional image to the training data set of the first detection model and add the second two-dimensional image to the training data set of the second detection model.

[0060] Exemplarily, since the third detection result is the detection result after manual analysis and has the highest credibility, the first two-dimensional image and the second two-dimensional image are annotated according to the third detection result. The annotated first two-dimensional image and the second two-dimensional image can be used as training samples, and the first two-dimensional image is added to the training data set of the first detection model and the second two-dimensional image is added to the training data set of the second detection model, and feedback is given to the first detection model and the second detection model to correct the first detection model and the second detection model, such as retraining the model or adjusting the network parameters to improve the detection capability of the detection model.

[0061] Through the above solution, a first industrial CT scanning device is used in combination with a first detection model, a second industrial CT scanning device is used in combination with a second detection model, and a third industrial CT scanning device is used in combination with manual analysis to achieve hierarchical detection of target products. The first industrial CT scanning device and the first detection model are used to perform a first-level detection to quickly detect defective products and improve overall detection efficiency. For target products whose first detection results are characterized as pending, a second or third detection is performed as appropriate, integrating the advantages of different CT detection methods to achieve efficient and accurate detection of products. The two-dimensional image is annotated based on the detection results and added to the training data set of the detection model to optimize the detection model.

[0062] In some application examples, the first detection model includes a first detection module and a first classification module. The first detection module is constructed based on the YOLOv11 model, and the size of the anchor box of the YOLOv11 model is configured to be multiple of 16×16, 32×32, 64×64 and 128×128 pixels. The first classification module is constructed based on the EfficientNet-B model, and the number of input channels of the EfficientNet-B model is configured to be a single channel.

[0063] For example, refer to Figure 2 , Figure 2 The principle block diagram of the first detection model is shown in FIG. Figure 2 The left side of the figure shows the block diagram of the first detection module. The captured first two-dimensional image (i.e., the projection of the target product) is input to the input layer and normalized (also known as normalization) before being fed into the first detection module. The first detection module processes the input image sequentially through a downsampling layer, a multi-layer convolutional layer (Conv2D), a batch normalization layer (BatchNorm), an activation layer, an upsampling layer, and a regression layer + a classification layer. The number of convolutional layers can be adjusted based on the application, and the activation layer uses the LeakyReLU method. The right side of the figure shows the block diagram of the first classification model, which is based on the EfficientNet-B model, such as any of the EfficientNet-B0-7 models. The feature map output by the upsampling layer of the first detection module is transmitted to the convolution layer of the first classification module. The convolution layer of the first classification module adopts depthwise separable convolution (DepthwiseConv2D) and two-dimensional convolution (Conv2D). The feature map processed by the convolution layer is sequentially transmitted to the Squeeze-and-Excitation layer (SE layer), the global average pooling layer (GlobalAveragePooling2D) and the fully connected layer (Dense), and activated by the Softmax function. Among them, the fully connected layer of the first classification module is connected to the regression layer + classification layer of the first detection module.

[0064] The first detection module is built based on YOLOv11. To address the small size of defects in industrial CT projection images, such as the average area of ​​defects ranging from 1% to 5%, the YOLOv11 model's anchor frame size is configured to multiple sizes ranging from 16×16 to 128×128 pixels, allowing it to cover small targets of multiple sizes. Multi-scale feature fusion technology is introduced through the upsampling layer, combining shallow detail features with deep semantic features to improve the ability to identify small defects. The first classification module uses a lightweight EfficientNet-B model, such as the EfficientNet-B0 model, and configures the input channel number to a single channel. While the conventional EfficientNet-B model has three channels to accommodate color images, this embodiment adjusts the channel number to a single channel to accommodate the grayscale features of CT projection images. This reduces the amount of computation while retaining most of the feature extraction capabilities, balancing efficiency and accuracy.

[0065] In some application examples, the first detection result includes a first defect location and a first confidence level. The first two-dimensional image is detected based on a pre-trained first detection model to obtain the first detection result, and then the following steps are further included:

[0066] When the first defect position is not empty or the first confidence level is lower than a first preset threshold, the first detection result is characterized as a pending result.

[0067] Exemplarily, a pre-trained first detection model detects a first two-dimensional image and outputs a first detection result including a first defect location and a first confidence level. In a specific example, the first detection result may include the first defect location, the first defect category, and the first confidence level. If the first defect location is not empty, it indicates that the first detection result may contain a defect. Alternatively, if the first confidence level is lower than a first preset threshold, such as 0.9, it indicates that the first detection result may contain a defect. The first detection result is then characterized as pending, allowing for further testing of the target product.

[0068] The second detection model includes a second detection module and a second classification module. The second detection module is built based on the enhanced YOLOv11 model. The enhanced YOLOv11 model uses a feature pyramid network to fuse shallow features and deep features. The shallow features are used to represent the feature layer with an image resolution of 1024×1024, and the deep features are used to represent the feature layer with an image resolution of 128×128 pixels. The second classification module is based on the EfficientNet-lite model.

[0069] For example, please refer to Figure 3 , Figure 3The principle block diagram of the second detection model is shown in the figure. The second two-dimensional image (such as a planar CT slice image) is input into the second detection module after preprocessing. The preprocessing includes noise reduction (MedianFilter2D), grayscale correction (HistogramEqualization) and normalization (Normalize). The second detection module is constructed based on the YOLOv11 model. The backbone part (Backbone) includes Conv2D (two-dimensional convolution) + MaxPooling2D (maximum pooling), Conv2D + BatchNorm (batch normalization) + SiLU (an activation function), and then passes through the depth-wise separable convolution + attention layer, and then repeats the depth-wise separable convolution. The number of repetitions can be determined according to the actual application. The bottleneck (Neck) part fuses multi-scale features through the upsampling layer to generate a pyramid feature map, and then passes through the regression layer + classification layer to calculate the confidence of the output result. The second classification module is based on the EfficientNet-lite model, such as the EfficientNet-lite3 model. The feature extraction layer includes Conv2D+BatchNorm+SiLU. After processing by multiple MBConv modules, the final feature map is output and then enters the classification layer. The classification layer uses global average pooling (GlobalAveragePooling2D).

[0070] It is worth mentioning that the second detection module of this embodiment is based on the enhanced YOLOv11 model and uses the feature pyramid network to fuse shallow features and deep features to improve the detection capability of defects of different scales (such as 5μm~500μm) and improve the overall detection accuracy of cross-scale defects.

[0071] In some application examples, the EfficientNet-lite model's MBConv module incorporates a Squeeze-and-Excitation (SE) module, such as a channel dilation + SE module + Dropout (random dropout). While the conventional EfficientNet-Lite model removes the SE module for lightweight performance, this embodiment re-embeds a customized SE module within the EfficientNet-Lite model. This adapts the input to a single channel for texture differentiation in single-channel grayscale planar CT images, using a channel-attention mechanism to enhance feature weighting for defects with similar grayscale textures, such as cracks and delamination. Compared to the conventional EfficientNet-B model, this SE module improves the classification accuracy of similar defects by restoring channel attention while maintaining a lightweight architecture (adapting to the computational efficiency of high-resolution images), achieving a balance between lightweight performance and fine-grained differentiation.

[0072] In some application examples, the second detection model is configured with a reinforcement loss function during the training phase. The calculation formula of the reinforcement loss function is:

[0073] ,

[0074] Where L represents the detection loss of the second two-dimensional image, α δ Characterized as the defect scale weighting factor, L small Characterized by the positioning loss of small size defects. Small size defects are used to characterize defects with an area less than 100μm. 2 The defect scale weighting factor can be dynamically adjusted according to the defect area ratio, and the adjustment range is 0.8~1.5, which can focus on optimizing the detection accuracy of tiny defects.

[0075] The calculation formula for the positioning loss of small-size defects is:

[0076] ,

[0077] In the formula, IOU is used to represent the intersection-over-union ratio between the predicted box A and the real box B, and C is used to represent the minimum enclosing rectangle of the predicted box A and the real box B, which can solve the failure problem of IOU being 0 when positioning small targets.

[0078] The following describes in detail the industrial CT hierarchical detection method according to an embodiment of the present invention using a specific application example.

[0079] 1. Level 1 testing

[0080] Data Acquisition: Utilizing high-precision industrial CT scanning equipment, the target product is comprehensively scanned to obtain a first two-dimensional image. During the scanning process, scanning parameters such as tube voltage, tube current, and scan time are appropriately adjusted based on the target product's material, shape, and the expected defect type to obtain clear and complete X-ray projection images. After the scan is complete, the raw projection image data is preprocessed, such as by removing noise and correcting the image's grayscale values, to improve image quality.

[0081] Training the first detection model: The first detection model includes a first detection module and a first classification module. The first detection module is built based on the YOLOv11 model, and the first classification module is built based on the lightweight EfficientNet-B0 model. In this way, the fusion architecture based on the YOLOv11 model and the EfficientNet-B0 model can be used as the core network for projection image detection. The specific training process is as follows:

[0082] Constructing a training dataset: We collected industrial CT projection image samples containing different materials and defect types, including metal and non-metal materials, and defect types such as cracks and poor solder joints. The sample size exceeded 100,000, of which 30% were defective. The data covered imaging data under multiple scanning parameters, including tube voltages of 80 to 220 kV and tube currents of 5 to 50 mA. We performed pixel-level annotation of defect areas in the sample images, generating a label file containing defect coordinates, category, and severity rating to construct a structured training dataset.

[0083] Data preprocessing: Data augmentation techniques were used to expand the sample dataset to more than 500,000 samples. These techniques included rotation, adding Gaussian noise, and histogram equalization. Grayscale normalization was performed on the sample images to ensure a standardized grayscale distribution for the input network.

[0084] Design of network architecture: The first detection module is built based on the YOLOv11 model and introduces multi-scale feature fusion, such as the feature maps of Conv4_3 and Conv5_3, where Conv4_3 identifies the third convolutional layer of the fourth Conv block. The anchor box sizes are optimized to 16×16, 32×32, 64×64, and 128×128 pixels based on the defect scale distribution of industrial CT projection images, for example, an average area ratio of 1% to 5%, to improve small target detection performance. The first classification module is based on the lightweight EfficientNet-B0 model and adjusts the number of input channels from 3 to 1 channel. This helps adapt to the grayscale characteristics of CT projection images and takes into account both computational efficiency and feature extraction capabilities.

[0085] Design loss function: Construct a multi-task joint loss function, including target localization loss function, category confidence loss function and classification loss function;

[0086] Among them, the multi-task joint loss function L=λ coord L coord +λ conf L conf +λ class L class , where λ coord ,λ con ,λ class They are the target positioning loss function L coord , category confidence loss function L conf And the classification loss function L class The target positioning loss function is calculated using mean square error, the category confidence loss function is calculated using binary cross entropy, and the classification loss function is calculated using cross entropy.

[0087] Training optimization: Use Adam optimizer with initial learning rate 10 -4, using a cosine annealing decay strategy; batch size 32, training cycle 500 epochs; calculate mAP (mean average precision), F1-score (F1 score) and other indicators on the validation set (accounting for 20%) every 5 epochs, triggering early stopping when the mAP improvement is less than 0.1% for 20 consecutive epochs; non-maximum suppression (NMS) is performed on the detection results with a threshold set of 0.5 to ensure that only the highest confidence prediction box is retained in the single defect area.

[0088] The pre-trained first detection model can accurately identify defect information in the projection image. During the detection process, the first detection model outputs the first defect location, first defect category, and first confidence level for each potential defect area. A confidence threshold is set, such as 0.9, and the first detection result is determined based on the first defect location, first defect category, and first confidence level, achieving rapid, one-shot detection.

[0089] 2. Secondary testing

[0090] Sample screening: For target products marked as defective or with low confidence in the first-level inspection of the projection image, especially individual parts such as connectors and chips, planar CT inspection is performed on the target products.

[0091] Planar CT Scanning: The screened target product is placed on a planar CT scanner. Leveraging the advantages of planar CT's high theoretical spatial resolution and the X-ray source's ability to closely track the target product, a detailed scan is performed. During the scanning process, parameters such as slice thickness and spacing are adjusted based on the target product's size, shape, and inspection requirements to obtain high-quality planar CT image data.

[0092] Training the second detection model: Targeting the high-resolution characteristics of planar CT images, such as single-pixel size ≤ 5μm, the second detection model includes a second detection module and a second classification module. The second detection module is built based on the enhanced YOLOv11 model, and the second classification module is built based on the EfficientNet-Lite3 model, resulting in a fusion architecture of the YOLOv11 and EfficientNet-Lite3 models. The training process is as follows:

[0093] Constructing a data set: Using planar CT scanning equipment to scan images of precision parts such as joints and chips, the images contain micron-level defects and are labeled with sub-pixel accuracy. In addition to conventional data enhancement operations, edge enhancement (Sobel operator) and local contrast normalization are added to highlight the edge features of tiny defects.

[0094] Network structure adaptation: The second detection module adds a feature pyramid network (FPN) based on YOLOv11, using the feature pyramid network to fuse shallow features with deep features. The input image of the second detection module has a resolution of 2048×2048 pixels, the resolution of the fused shallow features is 1024×1024 pixels, and the resolution of the deep features is 128×128 pixels. This helps improve the detection capabilities of defects of different scales, such as 5μm to 500μm. It also enhances the ability to distinguish similar defects such as cracks and delamination based on the grayscale texture features of planar CT images.

[0095] Loss function enhancement: The calculation formula of the enhanced loss function is:

[0096] ,

[0097] Where L represents the detection loss of the second two-dimensional image, α δ Characterized as the defect scale weighting factor, L small Characterized by the positioning loss of small size defects. Small size defects are used to characterize defects with an area less than 100μm. 2 The defect scale weighting factor can be dynamically adjusted according to the defect area ratio, and the adjustment range is 0.8~1.5, which can focus on optimizing the detection accuracy of tiny defects.

[0098] The calculation formula for the positioning loss of small-size defects is:

[0099] ,

[0100] In the formula, IOU is used to represent the intersection-over-union ratio between the predicted box A and the real box B, and C is used to represent the minimum enclosing rectangle of the predicted box A and the real box B, which can solve the failure problem of IOU being 0 when positioning small targets.

[0101] Adjust the training strategy: uniformly crop or scale images of different resolutions to a size suitable for the second detection model, such as 2048×2048 pixels, and use overlapping cropping, for example, with an overlap rate of 25%, to solve the memory usage problem of large-size images while preserving defect integrity. Reduce the learning rate to 5*10 -5 , extend the training cycle to 800 epochs, and use focal loss to alleviate the imbalance problem of positive and negative samples, for example, the proportion of positive samples is less than 5%.

[0102] The trained second inspection model has excellent defect detection capabilities. During inspection, the target product, having passed the primary inspection, is placed on a planar CT scanner. Leveraging the device's high theoretical spatial resolution and close proximity of the X-ray source to the sample, the product is scanned in detail. During the scanning process, parameters such as slice thickness and spacing are adjusted based on the product's size, shape, and inspection requirements to obtain high-quality planar CT image data, known as the second two-dimensional image. The pre-trained second inspection model analyzes and determines defects in the planar CT image data. Leveraging its high resolution, the second inspection model more accurately identifies tiny defects and defects in complex structures, and outputs a second inspection result, which includes a second defect location, second defect category, and second confidence level. The second inspection model's results are compared with those of the first inspection model. If the two results agree—that is, if both conclude that the target product has defects of the same type and location, or if both conclude that the product is defect-free—the target product is marked as good and the final inspection result is recorded. If the judgment results are inconsistent, for example, the first-level test shows a defect while the second-level test shows no defect, or the two have different judgments on the type and location of the defect, the target product will be marked as pending and enter the third-level test.

[0103] 3. Level 3 testing

[0104] Due to limited laboratory data, the first and second detection models may have imperfections in certain complex situations, leading to inconsistent judgment results. In this case, manual inspection with cone-beam CT is used for further precise detection.

[0105] Cone-beam CT scanning: The target product is placed on the cone-beam CT scanner. Leveraging its wide imaging range, the system performs a detailed, holistic scan of the target product. During the scanning process, the operator adjusts scanning parameters such as voltage, current, and scanning angle based on the product's characteristics and potential defects to achieve optimal imaging results. After the scan is complete, the acquired cone-beam CT image data is reconstructed into a 3D model of the target product's internal structure.

[0106] Manual evaluation: Professional inspectors use the 3D models and image data acquired through cone-beam CT scans, combined with their expertise and extensive testing experience, to perform manual evaluations of products. By analyzing the product's internal structural features, they identify potential defects and accurately determine their type, size, location, and severity, resulting in the most accurate test results.

[0107] Model Calibration: The results of manual cone-beam CT inspection are fed back to the first and second inspection models. By comparing the manual inspection results with the outputs of the first and second inspection models, any deviations and deficiencies in the inspection process of the first and second inspection models are identified. Based on the feedback, the first and second inspection models are then calibrated. For example, the first and second 2D images are annotated based on the third inspection results, and the first and second 2D images are added to the training dataset of the first and second inspection models, respectively. By retraining the models, adjusting network parameters, or adding more training samples, the inspection model structures are optimized, improving the models' detection capabilities and enabling more accurate detection of product defects.

[0108] By differentiating between tiered testing methods and integrating the application scenarios of planar CT and cone-beam CT, we have established a highly efficient testing system characterized by "efficient initial screening, precise final testing, and dynamic calibration." In practical applications, we continuously collect and analyze test data, continuously optimize the tiered testing model and parameters, and further improve testing efficiency and accuracy.

[0109] Please refer to Figure 4 , an embodiment of the present invention further provides an industrial CT hierarchical detection device, comprising:

[0110] A first acquisition module 110 is configured to acquire a first two-dimensional image of a target product provided by a first industrial CT scanning device;

[0111] A first determination module 120 is configured to detect the first two-dimensional image based on a pre-trained first detection model to obtain a first detection result;

[0112] A second acquisition module 130 is configured to acquire, when the first detection result indicates that the result is pending, a second two-dimensional image of the target product provided by a second industrial CT scanning device, wherein the second industrial CT scanning device has a higher theoretical spatial resolution than the first industrial CT scanning device;

[0113] A second determination module 140 is configured to detect the second two-dimensional image based on a pre-trained second detection model to obtain a second detection result;

[0114] A third acquisition module 150 is configured to acquire, when the second detection result is characterized as pending, a three-dimensional image model of the target product including internal structural features provided by a third industrial CT scanning device;

[0115] A third determination module 160 is configured to display the three-dimensional image model for manual analysis to obtain a third detection result;

[0116] Feedback module 170 is used to annotate the first two-dimensional image and the second two-dimensional image according to the third detection result, and add the first two-dimensional image to the training data set of the first detection model and add the second two-dimensional image to the training data set of the second detection model.

[0117] The inventive concept of the embodiment of the industrial CT grading detection device is consistent with the inventive concept of the embodiment of the industrial CT grading detection method described above. The contents and beneficial effects not covered in the embodiment of the industrial CT grading detection device can be referred to the embodiment of the industrial CT grading detection method described above and will not be repeated here.

[0118] Please refer to Figure 5 , an embodiment of the present invention also provides an industrial CT grading detection system, including a processor 210 and a memory 220, wherein a computer program is stored in the memory 220, and the processor 210 is used to implement the above-mentioned industrial CT grading detection method when running the computer program. Among them, the industrial CT grading detection method can be referred to above and will not be described in detail here. A first industrial CT scanning device is used in combination with a first detection model, a second industrial CT scanning device is used in combination with a second detection model, and a third industrial CT scanning device is used in combination with manual analysis to achieve graded detection of target products. A first-level detection is performed by the first industrial CT scanning device and the first detection model to quickly detect defective products and improve overall detection efficiency. For target products whose first detection results are characterized as pending results, a second or third detection is performed as appropriate, integrating the advantages of different CT detection methods to achieve efficient and accurate detection of products, and annotating the two-dimensional image based on the detection results and adding it to the training data set of the detection model to optimize the detection model.

[0119] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and when the computer program is run, the above-mentioned industrial CT hierarchical detection method is implemented. Among them, the industrial CT hierarchical detection method can be referred to above and will not be repeated here. A first industrial CT scanning device is used in combination with a first detection model, a second industrial CT scanning device is used in combination with a second detection model, and a third industrial CT scanning device is used in combination with manual analysis to achieve hierarchical detection of target products. A first-level detection is performed by the first industrial CT scanning device and the first detection model to quickly detect defective products and improve overall detection efficiency. For target products whose first detection results are characterized as pending results, a second or third detection is performed as appropriate, integrating the advantages of different CT detection methods to achieve efficient and accurate detection of products, and annotating the two-dimensional image based on the detection results and adding it to the training data set of the detection model to optimize the detection model.

[0120] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the scope of the present invention.

Claims

1. An industrial CT hierarchical detection method, characterized in that: include: Acquire a first two-dimensional image of a target product provided by a first industrial CT scanning device; Performing detection on the first two-dimensional image based on a pre-trained first detection model to obtain a first detection result, the first detection model including a first detection module and a first classification module, the first detection module being constructed based on a YOLOv11 model, the anchor box size of the YOLOv11 model being configured as multiple of 16×16, 32×32, 64×64, and 128×128 pixels, the first classification module being constructed based on an EfficientNet-B model, the input channel number of the EfficientNet-B model being configured as a single channel, the upsampling layer of the first detection module being connected to the convolutional layer of the first classification module, and the fully connected layer of the first classification module being connected to a module consisting of a regression layer and a classification layer of the first detection module; When the first detection result is characterized as pending, obtaining a second two-dimensional image of the target product provided by a second industrial CT scanning device, the second industrial CT scanning device having a higher theoretical spatial resolution than the first industrial CT scanning device; The second two-dimensional image is detected based on a pre-trained second detection model to obtain a second detection result, where the second detection model includes a second detection module and a second classification module. The second detection module is constructed based on an enhanced YOLOv11 model. The enhanced YOLOv11 model uses a feature pyramid network to fuse shallow features and deep features. The shallow features are used to represent a feature layer with an image resolution of 1024×1024 pixels, and the deep features are used to represent a feature layer with an image resolution of 128×128 pixels. The second classification module is based on an EfficientNet-lite model, and an SE module is added to the MBConv module of the EfficientNet-lite model. When the second detection result indicates that the result is pending, obtaining a three-dimensional image model of the target product including internal structural features provided by a third industrial CT scanning device; displaying the three-dimensional image model for manual analysis to obtain a third detection result; The first two-dimensional image and the second two-dimensional image are annotated according to the third detection result, and the first two-dimensional image is added to the training data set of the first detection model and the second two-dimensional image is added to the training data set of the second detection model.

2. The industrial CT hierarchical detection method according to claim 1, characterized in that: The first detection result includes a first defect position and a first confidence level. The first two-dimensional image is detected based on the pre-trained first detection model to obtain the first detection result. The method further includes: When the first defect position is not empty or the first confidence level is lower than a first preset threshold, the first detection result is characterized as pending.

3. The industrial CT hierarchical detection method according to claim 1, characterized in that: The second detection model is configured with an enhanced loss function during the training phase. The calculation formula of the enhanced loss function is: ， Where L represents the detection loss of the second two-dimensional image, α δ Characterized as the defect scale weighting factor, L small Characterized as positioning loss of small size defects, where the small size defect is used to characterize the defect area less than 100μm 2 defects.

4. The industrial CT hierarchical detection method according to claim 3, characterized in that: The calculation formula for the positioning loss of the small size defect is: ， In the formula, IOU is used to represent the intersection-over-union ratio between the predicted box A and the real box B, and C is used to represent the minimum enclosing rectangle of the predicted box A and the real box B.

5. An industrial CT hierarchical detection device, characterized in that: include: A first acquisition module is used to acquire a first two-dimensional image of a target product provided by a first industrial CT scanning device; a first determination module, configured to detect the first two-dimensional image based on a pre-trained first detection model to obtain a first detection result, wherein the first detection model includes a first detection module and a first classification module, the first detection module being constructed based on a YOLOv11 model, wherein the size of the anchor box of the YOLOv11 model is configured to be multiple of 16×16, 32×32, 64×64, and 128×128 pixels, the first classification module being constructed based on an EfficientNet-B model, wherein the number of input channels of the EfficientNet-B model is configured to be a single channel, the upsampling layer of the first detection module being connected to the convolutional layer of the first classification module, and the fully connected layer of the first classification module being connected to a module consisting of a regression layer and a classification layer of the first detection module; a second acquisition module, configured to, when the first detection result is characterized as pending, acquire a second two-dimensional image of the target product provided by a second industrial CT scanning device, the second industrial CT scanning device having a higher theoretical spatial resolution than the first industrial CT scanning device; a second determination module, configured to detect the second two-dimensional image based on a pre-trained second detection model to obtain a second detection result, the second detection model including a second detection module and a second classification module, the second detection module being constructed based on an enhanced YOLOv11 model, the enhanced YOLOv11 model fusing shallow features and deep features using a feature pyramid network, the shallow features being used to characterize a feature layer having an image resolution of 1024×1024 pixels, and the deep features being used to characterize a feature layer having an image resolution of 128×128 pixels, the second classification module being based on an EfficientNet-lite model, and an SE module being added to the MBConv module of the EfficientNet-lite model; a third acquisition module, configured to acquire, when the second detection result is characterized as pending, a three-dimensional image model of the target product including internal structural features provided by a third industrial CT scanning device; a third determination module, configured to display the three-dimensional image model for manual analysis to obtain a third detection result; A feedback module is used to annotate the first two-dimensional image and the second two-dimensional image according to the third detection result, and to add the first two-dimensional image to the training data set of the first detection model and to add the second two-dimensional image to the training data set of the second detection model.

6. An industrial CT hierarchical detection system, comprising a processor and a memory, wherein the memory stores a computer program, characterized in that: When the processor runs the computer program, it is used to implement the industrial CT hierarchical detection method according to any one of claims 1 to 4.

7. A storage medium storing a computer program, wherein: When the computer program is executed, the industrial CT hierarchical detection method according to any one of claims 1 to 4 is implemented.

Images