Part defect identification method and device based on 3D imaging and storage medium

Abstract

The invention discloses a part defect identification method and device based on 3D imaging and a storage medium, and belongs to the technical field of industrial quality detection. The method comprises the following steps: obtaining surface image data, obtaining an internal imaging data set, processing the surface image data set and the internal imaging data set, and outputting defect information. According to the method, the geometric feature information of the parts is extracted from the surface image data set to accurately plan the internal scanning path, and internal detection with high pertinence and high efficiency is realized.

Description

Technical Field

[0001] The present invention relates to the field of industrial quality inspection technology, and in particular to the field of component defect inspection. Background Art

[0002] As modern manufacturing continues to increase its requirements for product quality, component defect detection technology plays a vital role in ensuring product reliability and safety. Existing component detection technologies usually use a single surface detection or internal detection method, which has significant shortcomings in internal defect detection. Traditional internal scanning methods lack targeted planning and often use fixed scanning paths and parameters. They are unable to adaptively adjust according to the specific geometric features of parts, resulting in low scanning efficiency and easy omission of key areas. At the same time, existing technologies lack an effective fusion mechanism when processing surface and internal detection information. Surface image data and internal scanning data are analyzed independently, and the complementary advantages of the two cannot be fully utilized, affecting the overall detection accuracy and reliability. In addition, traditional anomaly detection methods rely on a large amount of labeled data, and face the challenges of difficult data acquisition and insufficient generalization capabilities in practical applications.

[0003] This technical solution proposes an intelligent internal inspection method guided by surface geometric feature information. By extracting component geometric feature information from the surface image dataset to accurately plan the internal scanning path, it achieves highly targeted and efficient internal inspection. This solution establishes a precise registration relationship between the surface inspection coordinate system and the internal inspection coordinate system, deeply fuses the surface image dataset with the internal imaging dataset, and forms fused image data containing complete structural information. By introducing an anomaly detection network built based on a weak supervision mechanism, this solution can achieve high-precision defect recognition under limited annotated data conditions, significantly improving the practicality and accuracy of the inspection system, and providing a more comprehensive and efficient technical solution for component quality control. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is that traditional internal scanning methods lack targeted planning, often use fixed scanning paths and parameters, and cannot be adaptively adjusted according to the specific geometric features of parts, resulting in low scanning efficiency and easy omission of key areas. In this regard, the embodiments of the present invention provide a component defect recognition method and device based on 3D imaging, and a storage medium, which accurately plans the internal scanning path by extracting component geometric feature information from the surface image data set, thereby achieving highly targeted and efficient internal inspection.

[0005] To solve the above technical problems, embodiments of the present invention provide a component defect recognition method, device, and readable storage medium based on 3D imaging. The method includes the following steps: S100: Collecting original image data of the component surface and establishing a surface detection coordinate system; performing image optimization processing on the original image data to obtain surface image data; performing spatial registration on the surface image data through the surface detection coordinate system to obtain a surface image data set; S200: Extracting component geometric feature information from the surface image dataset, establishing an internal detection coordinate system aligned with the surface detection coordinate system, acquiring component internal scanning data, planning the component internal scanning path based on the geometric feature information, imaging the internal scanning data using an imaging algorithm to obtain internal imaging data, and registering the internal imaging data to the internal detection coordinate system to obtain an internal imaging dataset; S300: Perform registration processing on the surface image dataset and the internal imaging dataset to generate fused image data, and extract candidate regions of interest from the fused image data through image preprocessing; divide the candidate regions of interest into image blocks and input them into the anomaly detection network constructed based on the weak supervision mechanism; perform feature processing and reconstruction on the image blocks through the anomaly detection network, calculate the reconstruction error and generate an anomaly score map; process the anomaly score map to extract the defect area and output the defect information.

[0006] In step S100, a structured light 3D imaging system is used to collect original image data of the component surface in multiple angles and time sequences by using structured light fringe projection profile and diffraction optical phase encoding.

[0007] The image optimization processing in step S100 includes: performing denoising filtering, brightness equalization and edge enhancement processing on the original image data to generate high-quality surface image data.

[0008] In step S200, the outer contour, thickness distribution and material boundary information of the component are extracted from the surface image data set as geometric feature information, and an internal detection coordinate system is established that is aligned with the surface detection coordinate system.

[0009] The acquisition of internal scanning data of the component in step S200 is achieved by laser ultrasonic scanning of the internal structure of the component.

[0010] In step S200, planning the internal scanning path of the component based on the geometric feature information includes: establishing a relationship model between the laser ultrasonic incident angle and the sound field excitation effect; and optimizing the scanning path using a genetic algorithm to generate an optimal scanning path including the coordinate position and the laser ultrasonic incident angle.

[0011] In step S300 , the image data is registered to generate fused image data, specifically by performing geometric registration and scale normalization on the surface image dataset and the internal image dataset of the component to generate fused image data in a unified coordinate system.

[0012] In step S300, feature processing and reconstruction are performed on the image block through the network, and the reconstruction error is calculated to generate an anomaly score map. Specifically, the anomaly detection network constructed based on the weak supervision mechanism includes an encoder and a decoder. The encoder extracts features from the original image block to obtain a feature vector, and the decoder reconstructs the image block based on the feature vector to obtain a reconstructed image block; by calculating the pixel-level reconstruction error between the original image block and the reconstructed image block, the mean square error or structural similarity index is used to determine the degree of abnormality of each pixel point, and the degree of abnormality is mapped to a score value to generate an anomaly score map.

[0013] A device for component defect recognition based on 3D imaging, comprising a surface image data set collection module, an internal imaging data set collection module, and a defect information output module. The surface image data set collection module is used to collect original image data of the component surface and establish a surface detection coordinate system; perform image optimization processing on the original image data to obtain surface image data, and perform spatial registration on the surface image data through the surface detection coordinate system to obtain a surface image data set; the internal imaging data set collection module is used to extract component geometric feature information from the surface image data set, establish an internal detection coordinate system registered with the surface detection coordinate system, and collect internal scanning data of the component. , plan the internal scanning path of the component based on the geometric feature information, image the internal scanning data through the imaging algorithm to obtain the internal imaging data, align the internal imaging data to the internal detection coordinate system, and obtain the internal imaging data set; the defect information output module is used to perform registration processing on the surface image data set and the internal imaging data set to generate fused image data, and extract the candidate region of interest of the fused image data through image preprocessing; divide the candidate region of interest into image blocks and input them into the anomaly detection network constructed based on the weak supervision mechanism, perform feature processing and reconstruction on the image blocks through the anomaly detection network, calculate the reconstruction error to generate the anomaly score map; process the anomaly score map to extract the defect area and output the defect information.

[0014] A readable storage medium, characterized in that a program or instruction is stored on the readable storage medium, and when the program or instruction is executed by a processor, a step such as any one of steps s100-s300 of a component defect recognition method based on 3D imaging is implemented.

[0015] The present invention achieves all-round detection of surface and internal defects by adopting a structured light 3D imaging system combined with laser ultrasonic scanning technology. Specifically, the multi-angle and multi-time series acquisition method of structured light stripe projection profile and diffraction optical phase encoding effectively eliminates the blind spot problem existing in traditional single-view imaging and significantly improves the detection rate of surface defects. By extracting geometric feature information such as the outer contour, thickness distribution and material boundary of the component to establish an internal detection coordinate system, it provides an accurate spatial positioning basis for laser ultrasonic scanning. The scanning path optimization based on genetic algorithm further improves the detection efficiency. By establishing a relationship model between the laser ultrasonic incident angle and the sound field excitation effect, and using the sound field excitation effect, path length and detection time as a multi-objective optimization function, the detection path planning is made more reasonable. Compared with the traditional uniform scanning method, the detection time can be shortened, while ensuring the detection accuracy of key areas. The denoising filter, brightness equalization and multi-view image fusion algorithm in the image optimization processing effectively solve common problems in practical applications such as uneven illumination and shadow occlusion, providing a high-quality data foundation for subsequent defect identification.

[0016] The present invention adopts a variational autoencoder anomaly detection network based on a weak supervision mechanism, which overcomes the limitation of traditional supervised learning that requires a large amount of labeled data, and has significant advantages in practical engineering applications where defect samples are scarce. Through the intelligent extraction of candidate regions of interest, combined with the edge detection algorithm to identify geometric discontinuity areas and internal density change areas, the detection range is effectively narrowed, the computational complexity is reduced by more than 60%, and the detection efficiency is significantly improved. The variational autoencoder architecture can learn the data distribution characteristics of normal parts through the encoder-latent space-decoder design, and has a higher sensitivity to abnormal areas. The reconstruction error calculation is combined with the abnormal score map generated by the latent space distribution characteristics, so that the defect positioning accuracy reaches the sub-millimeter level. The post-processing of the abnormal score map can not only accurately extract the defect area, but also automatically classify the defect type and evaluate the severity level by setting adaptive thresholds, connected domain analysis and geometric parameter calculations, providing a quantitative basis for quality control and maintenance decisions. DETAILED DESCRIPTION

[0017] The following will be combined with the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0018] The terms "first," "second," and the like in the specification and claims of this application are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of this application can be implemented in an order other than that illustrated or described herein, and that the objects distinguished by "first," "second," and the like are generally of the same type, and do not limit the number of objects; for example, the first object can be one or more. In addition, the term "and / or" in the specification and claims refers to at least one of the connected objects, and the character " / " generally indicates that the objects connected are in an "or" relationship.

[0019] S100: Collecting original image data of the component surface and establishing a surface detection coordinate system; performing image optimization processing on the original image data to obtain surface image data; performing spatial registration on the surface image data through the surface detection coordinate system to obtain a surface image data set.

[0020] Preferably, a structured light 3D imaging system is used to capture raw image data of the component's surface from multiple angles and in multiple time sequences, using structured light fringe projection profiles and diffraction optical phase encoding. Imaging system parameters are set as follows: fringe period 8-32 pixels, projection angle 15°-45°, phase encoding using a 4-step or 8-step phase shift algorithm, and multi-view acquisition from 6-12 observation positions within a 360° range around the component.

[0021] Specifically, a right-handed Cartesian coordinate system is established as the surface detection coordinate system. By setting the reference point P0 (x0, y0, z0), the image data collected from multiple perspectives are uniformly registered according to the coordinate transformation matrix T:

[0022] in is the rotation matrix, is the translation vector.

[0023] The registration accuracy is evaluated by the reprojection error:

[0024] E is the average reprojection error (in pixels), N is the total number of feature points, pᵢ is the observed image point coordinate (2D), Pᵢ is the corresponding 3D spatial point, proj() is the camera projection function, which projects the 3D point onto the 2D image plane. Threshold: E < 0.5 pixels indicates high registration accuracy. This processing method not only improves the spatial consistency of the data but also provides an accurate geometric reference for subsequent feature extraction and registration analysis.

[0025] Furthermore, the image optimization processing includes: performing denoising filtering, brightness equalization and edge enhancement processing on the original image data to generate high-quality surface image data.

[0026] Preferably, the edge enhancement process uses the Sobel operator to calculate the gradient, and the gradient amplitude formula is:

[0027] Where: |G|: gradient magnitude, indicating the edge strength of the image at that pixel; Gx: gradient component in the x-direction (horizontal direction), used to detect vertical edges; Gy: gradient component in the y-direction (vertical direction), used to detect horizontal edges; gradient direction: θ = arctan(Gy / Gx).

[0028] The enhanced image is:

[0029] The enhancement coefficient β is preferably set to 0.3. is the enhanced image pixel value, : Original image pixel value, β: Enhancement coefficient, controls the strength of edge enhancement, β = 0: no enhancement, β > 0: enhanced edge.

[0030] Furthermore, the optimized surface image data is spatially registered through the surface detection coordinate system to obtain a surface image dataset. The registration process uses the ICP algorithm for precise alignment, and the objective function is

[0031] Where E is the total registration error (least squares error), N is the total number of corresponding point pairs, pᵢ is the i-th point in the target point cloud (3D coordinates), qᵢ is the i-th point in the source point cloud (3D coordinates), and T is a 4×4 homogeneous transformation matrix consisting of a rotation matrix R (3×3) and a translation vector t (3×1).

[0032] The iterative convergence condition is set to the mean square error change less than 10⁻ 6 The maximum number of iterations is 100. The final surface image dataset contains 3D point cloud coordinates, surface normal vectors, RGB color information, reflection intensity values, and texture feature vectors, providing a high-quality data foundation for subsequent processing.

[0033] S200: Extract the component's geometric feature information from the surface image data set, establish an internal detection coordinate system aligned with the surface detection coordinate system, collect component internal scanning data, plan the component's internal scanning path based on the geometric feature information, image the internal scanning data through an imaging algorithm to obtain internal imaging data, align the internal imaging data to the internal detection coordinate system, and obtain an internal imaging data set.

[0034] Specifically, the component's outer contour, thickness distribution, and material boundary information are extracted from the surface image dataset as the geometric feature information. Outer contour extraction utilizes the Canny edge detection algorithm, with detection parameters set to a low threshold of 50 and a high threshold of 150. Thickness distribution is obtained by calculating a minimum bounding box from the surface point cloud data. Material boundary information is identified by differences in reflection intensity and changes in texture features. An internal detection coordinate system is established, which is aligned with the surface detection coordinate system.

[0035] Preferably, the internal structure of the component is scanned by laser ultrasound to collect internal scanning data of the component. The laser ultrasound system parameters are set to: laser power density 1-10 MW / cm², pulse width 10-100 ns, repetition frequency 1-100 Hz, ultrasonic detection frequency range 0.1-20 MHz, and sampling rate not less than 100 MS / s.

[0036] Specifically, the internal scanning path of the component is planned based on the geometric feature information. A relationship model between the laser ultrasonic incident angle θ and the acoustic field excitation effect η is established:

[0037] in is the acoustic field excitation effect at the incident angle θ, is the maximum excitation effect at vertical incidence, preferably 1.0, is the angle attenuation parameter, preferably 0.02 rad⁻², is the laser incident angle in radians.

[0038] Furthermore, a genetic algorithm is used to optimize the scanning path, and the objective function is set as:

[0039] Where F is the path optimization evaluation coefficient, the larger the value, the better the path quality. is the sound field excitation effect of the i-th point, L is the total path length in mm, T is the total detection time in s, L0 is the reference path length, preferably the shortest path length, T0 is the reference value, preferably the theoretical minimum detection time, α, β, γ are weight coefficients, preferably 0.6, 0.2, 0.2 respectively.

[0040] The individual coding of the genetic algorithm adopts real number coding, and each individual is represented as:

[0041] in, , is the parameter combination of the i-th scanning point, ( ) is the three-dimensional coordinate of the i-th scanning point is the laser incident angle at the i-th scanning point, and its value range is [0°, 60°].

[0042] Preferably, the internal scanning data is imaged by an imaging algorithm to obtain internal imaging data; synthetic aperture focusing technology (SAFT) is used for ultrasonic imaging, and the time domain SAFT algorithm is:

[0043] Where I(x,z) is the imaging intensity at the coordinate (x,z), is the total number of transducers, t is the time delay, xᵢ is the transducer position, c is the speed of sound, is the propagation time of the sound wave from emission to reception.

[0044] The frequency domain beamforming algorithm is

[0045] Where S(ω) is the received signal in the frequency domain, obtained by Fourier transform, is the angular frequency, is the phase compensation factor used for focused imaging. The imaging resolution is set to 0.1 mm in the axial direction and 0.2 mm in the lateral direction, and the dynamic range is not less than 40 dB.

[0046] Furthermore, the internal imaging data is registered to the internal detection coordinate system to obtain an internal imaging data set.

[0047] S300: Perform registration processing on the surface image dataset and the internal imaging dataset to generate fused image data, and extract candidate regions of interest from the fused image data through image preprocessing; divide the candidate regions of interest into image blocks and input them into an anomaly detection network constructed based on a weak supervision mechanism, perform feature processing and reconstruction on the image blocks through the anomaly detection network, calculate the reconstruction error and generate an anomaly score map; process the anomaly score map to extract defect areas, and output defect information.

[0048] Specifically, the component's surface image dataset and internal imaging dataset are geometrically registered and scale-normalized to generate fused image data in a unified coordinate system. The registration process utilizes a feature point matching method, extracting key feature points from the surface and interior to establish a correspondence, with a registration error controlled within 0.02 mm. Scale normalization unifies data of varying resolutions to a standard voxel size of 0.1 mm × 0.1 mm × 0.1 mm.

[0049] Preferably, candidate regions of interest (ROIs) are extracted from the fused image data through image preprocessing. This preprocessing includes morphological operations and a region growing algorithm. The morphological opening kernel size is set to 5×5, and the region growing seed point threshold is set to the fused image mean plus 1.5 standard deviations. Candidate ROI screening criteria include an area greater than 100 pixels and an aspect ratio within the range of 0.2-5.0.

[0050] Specifically, the candidate regions of interest are divided into image blocks and input into an anomaly detection network built using a weakly supervised mechanism. The image blocks are set to 64×64 pixels in size, with an overlap ratio of 50% to ensure complete coverage. The anomaly detection network built using a weakly supervised mechanism includes an encoder and a decoder. The encoder extracts features from the original image blocks to generate feature vectors.

[0051] Furthermore, the encoder adopts a convolutional neural network structure, including 4 convolution layers, with the following parameters: the first convolution kernel is 3×3, stride=1, padding=1, and the number of output channels is 64; the second convolution kernel is 3×3, stride=2, padding=1, and the number of output channels is 128; the third convolution kernel is 3×3, stride=2, padding=1, and the number of output channels is 256; the fourth convolution kernel is 3×3, stride=2, padding=1, and the number of output channels is 512. The decoder reconstructs the image block based on the feature vector to obtain a reconstructed image block, adopts a transposed convolution structure symmetrical to the encoder, and the reconstruction loss function is:

[0052] Where I(i, j) is the pixel value of the original image block, Î(i, j) is the pixel value of the reconstructed image block, and H and W are the height and width of the image block, respectively.

[0053] Preferably, the pixel-level reconstruction error between the original image block and the reconstructed image block is calculated, the degree of abnormality of each pixel is determined using a mean square error or a structural similarity index, and the degree of abnormality is mapped to a score value to generate an abnormality score map.

[0054] Furthermore, the abnormality score map is processed to extract the defect area, and the defect information is output. Finally, the complete defect information including the defect location coordinates, size parameters, shape characteristics and severity classification is output, so as to realize the accurate identification and quantitative evaluation of the surface and internal defects of the parts.

[0055] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server or network device, etc.) to execute the methods of the various embodiments of the present application.

[0056] The above describes the embodiments of the present application, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which fall within the scope of protection of this application.

Claims

1. A component defect recognition method based on 3D imaging, characterized in that: The detection method comprises: S100: Collecting original image data of the component surface and establishing a surface detection coordinate system; performing image optimization processing on the original image data to obtain surface image data; performing spatial registration on the surface image data through the surface detection coordinate system to obtain a surface image data set; S200: extracting the component's geometric feature information from the surface image dataset, establishing an internal detection coordinate system registered with the surface detection coordinate system, acquiring component internal scanning data, planning a component internal scanning path based on the geometric feature information, imaging the internal scanning data using an imaging algorithm to obtain internal imaging data, and registering the internal imaging data to the internal detection coordinate system to obtain an internal imaging dataset; S300: Perform registration processing on the surface image dataset and the internal imaging dataset to generate fused image data, and extract candidate regions of interest from the fused image data through image preprocessing; divide the candidate regions of interest into image blocks and input them into an anomaly detection network constructed based on a weak supervision mechanism, perform feature processing and reconstruction on the image blocks through the anomaly detection network, calculate the reconstruction error and generate an anomaly score map; process the anomaly score map to extract defect areas, and output defect information.

2. The component defect recognition method based on 3D imaging according to claim 1, characterized in that: The step S100 uses a structured light 3D imaging system, structured light stripe projection profile and diffraction optical phase encoding to collect original image data of the surface of the component at multiple angles and in multiple time sequences.

3. The component defect recognition method based on 3D imaging according to claim 1, characterized in that: The image optimization processing in step S100 includes: performing denoising filtering, brightness equalization and edge enhancement processing on the original image data to generate the surface image data.

4. The component defect recognition method based on 3D imaging according to claim 1, characterized in that: In the step S200 , the outer contour, thickness distribution and material boundary information of the component are extracted from the surface image data set as the geometric feature information, and the internal detection coordinate system aligned with the surface detection coordinate system is established.

5. The component defect recognition method based on 3D imaging according to claim 4, characterized in that: The acquisition of the internal scanning data of the component in step S200 is achieved by laser ultrasonic scanning of the internal structure of the component.

6. The component defect recognition method based on 3D imaging according to claim 5, characterized in that: In the step S200, planning the internal scanning path of the component based on the geometric feature information includes: establishing a relationship model between the laser ultrasonic incident angle and the sound field excitation effect; and optimizing the scanning path using a genetic algorithm to generate the optimal scanning path including the coordinate position and the laser ultrasonic incident angle.

7. The component defect recognition method based on 3D imaging according to claim 1, characterized in that: In step S300, performing registration processing on the image data to generate fused image data specifically includes: performing geometric registration and scale normalization processing on the surface image dataset and the internal image dataset of the component to generate fused image data in a unified coordinate system.

8. The component defect recognition method based on 3D imaging according to claim 1, characterized in that: In step S300, the feature processing and reconstruction of the image block are performed through the network, and the reconstruction error is calculated to generate an abnormality score map. Specifically, the abnormality detection network constructed based on the weak supervision mechanism includes an encoder and a decoder, the encoder extracts features from the original image block to obtain a feature vector, and the decoder reconstructs the image block based on the feature vector to obtain a reconstructed image block; by calculating the pixel-level reconstruction error between the original image block and the reconstructed image block, the mean square error or structural similarity index is used to determine the abnormality degree of each pixel point, and the abnormality degree is mapped to a score value to generate an abnormality score map.

9. A device for component defect identification based on 3D imaging, comprising a surface image dataset collection module, an internal imaging dataset collection module, and a defect information output module, wherein the surface image dataset collection module is configured to collect raw image data of the component surface and establish a surface detection coordinate system; Performing image optimization processing on the original image data to obtain surface image data, and performing spatial registration on the surface image data through the surface detection coordinate system to obtain a surface image dataset; the internal imaging dataset collection module is used to extract the component geometric feature information from the surface image dataset, establish an internal detection coordinate system registered with the surface detection coordinate system, collect component internal scanning data, plan the component internal scanning path based on the geometric feature information, image the internal scanning data through an imaging algorithm to obtain internal imaging data, and register the internal imaging data to the internal detection coordinate system to obtain an internal imaging dataset; The defect information output module is used to perform registration processing on the surface image dataset and the internal imaging dataset to generate fused image data, extract candidate regions of interest from the fused image data through image preprocessing; divide the candidate regions of interest into image blocks and input them into an anomaly detection network constructed based on a weak supervision mechanism, perform feature processing and reconstruction on the image blocks through the anomaly detection network, calculate the reconstruction error to generate an anomaly score map; process the anomaly score map to extract the defect area, and output the defect information.

10. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the component defect identification method based on 3D imaging as described in any one of claims 1 to 8 are implemented.

Images