A multi-scene adaptive precision manufacturing high-precision nondestructive defect detection full-link system and method

Abstract

The application provides a multi-scene adaptive precision manufacturing high-precision nondestructive defect detection full-link system and method, containing a scheduling module, an FLI special detection model, a BLI special detection model and a magnification progressive trigger mechanism; wherein the scheduling module identifies different illumination modes of the to-be-detected image and distributes it to different detection models; the FLI special detection model detects surface defects; the BLI special detection model detects internal hole foreign matters; the magnification progressive trigger mechanism triggers the strategy of high-magnification precision detection when no effective defect is detected by the FLI special detection model and the BLI special detection model under low magnification. It is a complete engineering full-link technical system and can realize long-term stable detection in the production stage.

Description

Technical Field

[0001] This invention belongs to the fields of industrial machine vision, precision metrology and non-destructive testing technology, specifically involving defect detection technology in high-end precision manufacturing scenarios such as aerospace and semiconductors, and particularly involving a full-link integrated solution for front / back illumination multi-illumination mode adaptation, texture intensity hierarchical scheduling, post-inference security verification, and multi-magnification collaborative detection. Background Technology

[0002] In the field of high-end precision manufacturing, high-precision non-destructive testing of core process equipment such as precision cutting tools and micro-hole workpieces is a crucial link in ensuring product quality and avoiding losses in mass production. Existing mainstream testing solutions face three major systemic bottlenecks in industrial deployment:

[0003] 1. Texture interference leads to accuracy imbalance: The common use of single target detection model ignores the intensity difference of the inherent texture of the workpiece under front illumination (FLI). The false detection rate is high in the heavy texture area and the false detection rate of small defects in the light texture area is high. It is impossible to take into account the detection effect of both types of scenarios at the same time. Traditional texture smoothing preprocessing will erase the features of small defects, further reducing the recall rate.

[0004] 2. Mislabeling poses safety risks: Backlit (BLI) foreign object detection (FOD) scenarios lack targeted post-processing safety mechanisms, which can easily lead to mislabeling (FA) where the defect label falls entirely in a non-target area of ​​the workpiece. This can directly cause subsequent processing errors and result in irreversible damage to high-value workpieces.

[0005] 3. Poor adaptability to single-architecture scenarios: Existing solutions mostly focus on optimizing a single model structure and have not formed a full-link system of "scenario segmentation - accurate detection - post-inference verification - magnification adaptation", making it difficult to simultaneously meet the coordinated requirements of industrial inspection for accuracy, efficiency and long-term operational reliability; moreover, it is difficult to balance the detection of minor defects with production line cycle time, and independent modeling at different magnifications will significantly increase deployment and maintenance costs.

[0006] Currently, there is no solution that can systematically address all the above pain points. There is an urgent need to build a complete engineering end-to-end technology system to achieve long-term stable production-grade testing on a mature basic testing architecture. Summary of the Invention

[0007] The present invention aims to overcome the above-mentioned defects and propose a multi-scenario adaptive end-to-end high-precision non-destructive defect detection scheme.

[0008] This invention provides a multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing, characterized by: including a scheduling module, an FLI-specific detection model, a BLI-specific detection model, and a multiplier-progressive triggering mechanism;

[0009] The scheduling module identifies different lighting patterns in the image under test and assigns them to different detection models.

[0010] FLI-specific inspection model for detecting surface defects;

[0011] BLI-specific inspection model for detecting foreign objects in internal holes;

[0012] The progressive magnification triggering mechanism is a strategy that triggers high-magnification precision inspection when there are no defects. When neither the FLI-specific inspection model nor the BLI-specific inspection model detects a valid defect at low magnification, high-magnification inspection is triggered.

[0013] Furthermore, the multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing provided by the present invention is further characterized in that:

[0014] The FLI-specific detection model consists of a front-illuminated texture classification and scheduling submodule, a light texture-specific detection submodule, and a heavy texture-specific detection submodule.

[0015] Among them, the front-light texture classification and scheduling submodule classifies textures and distinguishes images into light texture images and heavy texture images;

[0016] A dedicated detection submodule for light textures is adapted to light texture surface scenarios and can detect minute defects.

[0017] A dedicated detection submodule for heavy textures is designed to detect and suppress false positives in heavy textured surface scenes.

[0018] Furthermore, the multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing provided by the present invention is further characterized in that:

[0019] Different thresholds are preset for both light-textured and heavy-textured surface scenes.

[0020] Furthermore, the multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing provided by the present invention is further characterized in that:

[0021] The BLI-specific detection model includes a BLI internal foreign object detection submodule and a BRV-GT safety filter submodule.

[0022] Among them, the BLI internal foreign object detection submodule identifies and locates foreign objects in the internal hole;

[0023] The BRV-GT safety filtering submodule uses a lightweight image processing algorithm to remove false positives from the BLI internal foreign object detection submodule.

[0024] Furthermore, the multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing provided by the present invention is further characterized in that:

[0025] The BRV-GT safety filtering submodule includes the following operating steps:

[0026] S31. Obtain the dynamic threshold:

[0027] The default method uses a weighted average of the mean and median to obtain the threshold.

[0028] When using a bimodal histogram device, a preset percentile gray value is used as the threshold.

[0029] S32. Noise reduction is achieved by using grayscale binarization and morphological opening operations to segment the effective area of ​​the inner hole from the non-target area of ​​the workpiece and remove small noise points;

[0030] S33. Check the grayscale distribution of each box and filter out any boxes that fall entirely in low grayscale non-target areas as invalid or incorrect annotations.

[0031] Furthermore, the multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing provided by the present invention is further characterized in that:

[0032] When the progressive multiplier trigger detection mechanism triggers high-multiplier detection, a targeted precision detection strategy is adopted;

[0033] Targeted precision inspection involves defining candidate regions based on the spatial distribution of historical defects and performing high-magnification precision scanning only on these candidate regions.

[0034] Furthermore, the multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing provided by the present invention is further characterized in that:

[0035] It also outputs the physical dimensions of the defect.

[0036] Furthermore, this invention also provides a multi-scenario adaptive end-to-end method for high-precision non-destructive defect detection in precision manufacturing, characterized by comprising the following steps:

[0037] S1. Acquire workpiece inspection images;

[0038] S2. Based on the illumination mode, assign the image to the corresponding detection path, i.e., the FLI detection path or the BLI detection path;

[0039] S3. Analyze the results via the FLI or BLI detection path and output the results;

[0040] When the result is "no valid defects detected", a high-magnification local fine inspection is triggered and the result is output.

[0041] Furthermore, the aforementioned multi-scenario adaptive high-precision non-destructive defect detection end-to-end method for precision manufacturing is further characterized by:

[0042] The FLI detection path includes the following steps:

[0043] S211. Classify the texture of the image;

[0044] S212. Based on different texture classification results, assign them to the corresponding detection paths, namely, light texture detection paths or heavy texture detection paths;

[0045] S213. Analyze the light texture detection path or the heavy texture detection path, and output the results based on the corresponding confidence threshold.

[0046] Furthermore, the aforementioned multi-scenario adaptive high-precision non-destructive defect detection end-to-end method for precision manufacturing is further characterized by:

[0047] The BLI detection path includes the following steps:

[0048] S221. A candidate frame for foreign objects in the internal hole is output by a dedicated detection model;

[0049] S222. Security verification is performed using the BRV-GT module to filter invalid / erroneous labels.

[0050] The function and effects of this invention:

[0051] Based on the solution of this invention, the problem of texture interference in FLI scenes is solved at the system level, while reducing the false detection rate of heavy texture scenes and the false detection rate of light texture scenes, and preserving the features of minor defects.

[0052] Based on the solution of this invention, mislabeling security filtering in BLI scenarios can be achieved with low deployment cost, completely eliminating the security risks of processing misoperation.

[0053] Based on the solution of this invention, a complete technical closed loop of "lighting subdivision - texture grading - dedicated detection - post-processing verification - magnification progression" is constructed, which takes into account detection accuracy, inference efficiency and engineering reliability.

[0054] Based on the solution of this invention, there is no need to rely on new basic testing models. It can meet production-level acceptance standards based on a mature architecture and is suitable for various high-end manufacturing scenarios such as aerospace and semiconductors. Attached Figure Description

[0055] Figure 1 This invention provides a full-chain overall architecture diagram of a multi-scenario adaptive high-precision non-destructive defect detection method and system for precision manufacturing;

[0056] Figure 2 Trend chart of online detection accuracy of production line over 57 days. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] This embodiment provides a multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing. Its system architecture includes a four-level adaptive end-to-end architecture consisting of a pre-processing texture classification and scheduling module, two FLI-specific detection models, a BLI-specific detection model, a grayscale verification security filtering module, and a set of progressive triggering mechanisms.

[0059] Based on this architecture, the system forms a closed-loop technology system: lighting and defect type segmentation → independent model detection → post-inference safety filtering → on-demand fine inspection based on magnification → iterative optimization of results. The specific solution is as follows:

[0060] S1. 3D Scene Division and Dataset Construction

[0061] The detection task is split into three dimensions: lighting mode, defect type, and texture intensity, and a task-specific dataset is constructed.

[0062] The illumination dimension includes front illumination (FLI, corresponding to surface defect detection) and back illumination (BLI, corresponding to internal hole foreign object detection).

[0063] The texture dimension includes light textures and heavy textures in FLI scenes;

[0064] The defect dimension includes surface defects and internal foreign matter defects.

[0065] In this embodiment, the dataset contains 9630 FLI texture scene images, including 8528 light textures and 1102 heavy textures; and 6463 BLI hole scene images.

[0066] For each image data in the dataset, three types of information are labeled: lighting type, texture intensity, and defect location bounding box.

[0067] The dataset is strictly divided into training and validation sets in a 95:5 ratio, with samples completely isolated and non-overlapping. Common data augmentation techniques such as random flipping, rotation, and brightness adjustment are used to improve the model's generalization ability.

[0068] Construction of the S2.FLI Surface Defect Adaptive Detection Subsystem

[0069] Composed of a front-light texture classification (FLTC) scheduling module and two dedicated object detection models, it achieves texture hierarchical adaptation:

[0070] The FLTC scheduling module employs a lightweight image classification architecture with an input resolution of 224×224. It classifies FLI images into two categories: light texture and heavy texture, which serve as the basis for subsequent model scheduling. This module achieves a classification accuracy of 98.70% on the independent validation set and takes approximately 24ms for single-image inference.

[0071] Dual dedicated detection models: both trained based on mature object detection architectures, with an input resolution of 640×640, including:

[0072] The Light Texture Detection Model (FLI-LTOD) is trained on the training set within the dataset, focusing on the detection of minute defects and adapted to light texture surface scenarios.

[0073] The FLI-HTOD (Fine Texture Detection Model) is trained on the training set within the dataset and focuses on suppressing false positives for textures, making it suitable for scenes with heavily textured surfaces.

[0074] Adaptive confidence threshold strategy: In this embodiment, the threshold for light texture scenes is set to 0.25 (prioritizing recall, which can be adjusted according to the actual needs of the scene), and the threshold for heavy texture scenes is set to 0.45 (prioritizing accuracy, which can be adjusted according to the actual needs of the scene). The texture is dynamically loaded based on the texture classification results.

[0075] Construction of the S3.BLI Internal Hole Inspection and Safety Verification Subsystem

[0076] Composed of the BLI internal foreign object detection model and the background region verification based on grayscale threshold (BRV-GT) module, it achieves a closed loop of accurate detection and safety protection. Specifically:

[0077] The BLI internal hole detection model is obtained by training on the training set within the dataset, with an input resolution of 640×640, and focuses on the accurate location of foreign objects in internal holes.

[0078] The BRV-GT security filtering module uses a lightweight image processing algorithm to remove false positive bounding boxes from the BLI internal hole detection model, retaining only the true defects and avoiding false alarms. It implements lightweight post-processing based on an open-source image processing library, with a single-frame verification overhead of only 3ms. The core logic is as follows:

[0079] S31. Obtain the dynamic threshold to locate the position range for subsequent repair stages:

[0080] The default is to use a weighted average of the mean and median (e.g., T = 0.6 × mean + 0.4 × median).

[0081] When using a bimodal histogram device, the 75th percentile gray value (which can also be adjusted according to actual needs) is used as the threshold; other threshold selection schemes can also be used to set the threshold.

[0082] S32. Denoise is removed by grayscale binarization and morphological opening operations, and the effective area of ​​the inner hole is separated from the non-target area of ​​the workpiece.

[0083] Specifically, the image is converted to black and white by grayscale binarization, and the image is divided into binary categories using grayscale thresholding to generate an inner hole mask, which quickly distinguishes between holes, cavities and the surrounding workpiece; this is used to determine whether the AI ​​detection box is within a reasonable detection range and to eliminate false defect alarms generated on the workpiece surface;

[0084] Noise reduction is achieved through opening operations, removing small noise points without compromising the true nature of the defects.

[0085] S33. Check the grayscale distribution of the region frame by frame. Annotation frames that fall completely in the low grayscale non-target area are judged as invalid erroneous annotations and filtered out.

[0086] S4. Multiplier Progression Trigger Detection Mechanism

[0087] A strategy combining hybrid magnification training and defect-free triggering for high-magnification precision inspection is adopted to balance the detection of minute defects with detection efficiency.

[0088] Training phase: Mixed training with 20× (can be any other low magnification) and 50× (can be any other high magnification) magnification data, without splitting independent models, reducing maintenance costs;

[0089] Triggering condition: When no valid defects are detected in FLI and BLI modes at 20× magnification, 50× high magnification detection is triggered. That is, if no problems are seen at low magnification, there may be very small defects that have been missed. At this time, high magnification fine inspection is started.

[0090] Targeted fine inspection: Based on the spatial distribution of historical defects, candidate areas are delineated, and only the candidate areas are scanned with 50× fine scan. Each trigger takes an additional 15-25ms.

[0091] Physical size calibration: 20× magnification pixel resolution 0.9375μm / pixel, 50× magnification 0.375μm / pixel, the detection results are directly output in micrometer level physical size, and the minimum detectable defect size of the system is 9μm.

[0092] The complete process based on the above system is as follows:

[0093] Acquire workpiece inspection images and assign them to the corresponding inspection subsystems according to the lighting mode;

[0094] FLI scenario: The texture intensity is determined by the FLTC module, and the corresponding dedicated detection model and confidence threshold are adaptively scheduled to output the candidate results of surface defects;

[0095] BLI scenario: The candidate bounding box for foreign objects in the internal hole is output by a dedicated detection model, and then the BRV-GT module is used to complete the safety verification and filter out invalid and erroneous annotations;

[0096] If no valid defects are detected in all 20x magnification modes, a 50x high-magnification local fine inspection is triggered.

[0097] It completes the conversion between pixels and physical dimensions, and outputs the final defect location, size, and type, realizing full-process non-destructive testing.

[0098] Based on the above method, this embodiment also provides corresponding system software / executable program, as follows:

[0099] The multi-scenario adaptive high-precision non-destructive defect detection full-link system for precision manufacturing provided in this embodiment includes a scheduling module, an FLI-specific detection model, a BLI-specific detection model, a multiplier progressive triggering mechanism, and a physical dimension adjustment module.

[0100] The scheduling module identifies different lighting patterns in the image under test and assigns them to different detection models.

[0101] The FLI-specific detection model detects surface defects and consists of a front-illuminated texture classification and scheduling submodule, a light texture-specific detection submodule, and a heavy texture-specific detection submodule. The front-illuminated texture classification and scheduling submodule classifies textures, distinguishing images into light texture images and heavy texture images. The light texture-specific detection submodule is adapted to light texture surface scenes and detects minute defects. The heavy texture-specific detection submodule is adapted to heavy texture surface scenes and detects and suppresses texture false positives.

[0102] The BLI-specific detection model detects foreign objects in internal orifices and includes a BLI internal orifice foreign object detection submodule and a BRV-GT safety filtering submodule. The BLI internal orifice foreign object detection submodule identifies and locates foreign objects in internal orifices. The BRV-GT safety filtering submodule uses a lightweight image processing algorithm to remove misjudged objects from the BLI internal orifice foreign object detection submodule.

[0103] The progressive magnification triggering mechanism is a strategy that triggers high-magnification precision inspection when there are no defects. When neither the FLI-specific inspection model nor the BLI-specific inspection model detects a valid defect at low magnification, high-magnification inspection is triggered.

[0104] The physical size adjustment module outputs the physical size of the defect based on the final detection magnification.

[0105] Based on the above methods and systems, the following effects were achieved:

[0106] 1. Improved system-level accuracy: The mAP@50 accuracy for FLI light texture scenes reached 95.27%, an improvement of 24.10 percentage points compared to the initial single model; the false positive rate for heavy texture scenes decreased by 72.3%; the mAP@50 accuracy for BLI internal hole detection reached 93.99%; the overall offline detection accuracy of the entire system was 94.50%, the peak comprehensive accuracy of the production line online was 96.33%, and the daily average accuracy was 92.33%, fully meeting the metrological acceptance standards of the precision manufacturing industry.

[0107] 2. Zero-risk security guarantee: The BRV-GT module has been verified through 57 days of continuous operation on the production line, with no effective defect frame errors and no invalid annotation omissions, completely eliminating the risk of damage to high-value workpieces caused by mislabeling; moreover, the module is lightweight and easy to deploy, requiring no additional computing power investment.

[0108] 3. Strong scene adaptability: Multi-dimensional scene segmentation and independent modeling can flexibly adapt to different lighting, texture and magnification conditions; end-to-end inference speed of 19-24FPS, meeting the real-time detection requirements of production lines.

[0109] 4. Low engineering implementation cost: Based on a mature detection architecture and open source toolchain, it does not rely on high-end hardware and new models. The model is easy to maintain and iterate, and can be quickly promoted to various high-end manufacturing scenarios such as aerospace, semiconductor, and automotive precision machining.

[0110] Application Examples

[0111] like Figure 1 As shown, this embodiment includes the following steps for non-destructive testing of precision manufacturing tools across all scenarios:

[0112] S1. Workpiece Image Acquisition Unit: The unit uses 20x optical magnification to acquire the original image of the workpiece and transmits the imaging data downstream. Specifically, it initially uses a 20× objective lens, is equipped with a vertically adjustable FLI light source and a bottom adjustable BLI light source, and acquires an image resolution of 640×480 pixels.

[0113] S2.20×FLI front-illuminated adaptive surface defect detection subsystem: Built-in FLTC texture classification and scheduling module, which switches the corresponding detection model according to the surface texture intensity of the workpiece, outputs surface defect candidate data, and outputs texture intensity results within 24ms;

[0114] S3. First-level judgment node:

[0115] The FLI-LTOD model is scheduled for light-textured scenes with a confidence threshold of 0.25, and the output is a surface defect detection box.

[0116] The FLI-HTOD model is scheduled for heavily textured scenes with a confidence threshold of 0.45, and the output results are suppressed after texture false detections.

[0117] Determine whether a valid defect is detected under the front light mode. If a defect is detected, proceed directly to the calibration unit; otherwise, proceed to the backlight detection process.

[0118] S4.20× Backlit BLI Internal Hole Detection and Safety Verification Subsystem: For the detection of hidden foreign objects in internal holes, the input image is processed by the BLI-specific detection model to output candidate boxes for foreign objects; the candidate boxes enter the BRV-GT module: after grayscale conversion, dynamic thresholds are calculated, morphological denoising is performed, and each box is verified one by one, filtering out invalid boxes that fall completely in the low grayscale background, and the verification is completed within 3ms.

[0119] S5. Second-level judgment node: Determine whether a valid defect is detected in backlight mode. If detected, it directly enters the calibration unit. If no valid defect is detected in FLI or BLI at 20×, it automatically switches to 50× objective lens for targeted fine inspection of historically high-incidence areas such as tool inner hole edges, with a corresponding pixel resolution of 0.375μm / pixel, and can detect foreign objects as small as 9μm.

[0120] S6.50× High Magnification Directional Fine Inspection Module: Based on historical defect distribution, it identifies high-incidence areas and performs fine inspection of minute defects through local high-magnification scanning;

[0121] S7. Result Fusion and Physical Size Calibration Unit: Unifies and summarizes multi-optical path and multi-magnification detection data, and completes pixel to micrometer size conversion;

[0122] S8. Final Inspection Output Unit: Standardized inspection results with unified output of defect coordinates, geometric dimensions, and defect classification.

[0123] like Figure 2 As shown, after 57 days of continuous operation on the production line, the peak comprehensive detection accuracy of this embodiment was 96.33%, the daily average accuracy was 92.33%, the end-to-end detection speed was 21 FPS, and there were no workpiece damage events caused by mislabeling.

Claims

1. A multi-scenario adaptive high-precision non-destructive defect detection end-to-end system for precision manufacturing, characterized in that: It includes a scheduling module, a dedicated FLI detection model, a dedicated BLI detection model, and a rate-incrementing triggering mechanism; The scheduling module identifies different lighting patterns in the image under test and assigns them to different detection models. The FLI-specific detection model is used to detect surface defects. The BLI-specific detection model is used to detect foreign objects in the internal bore. The aforementioned progressive triggering mechanism employs a strategy of triggering high-magnification precision inspection when no defects are detected. This strategy triggers high-magnification inspection when neither the FLI-specific inspection model nor the BLI-specific inspection model detects any valid defects at low magnification.

2. The multi-scenario adaptive high-precision non-destructive testing end-to-end system for precision manufacturing as described in claim 1, characterized in that: The FLI-specific detection model consists of a front-illuminated texture classification and scheduling submodule, a light texture-specific detection submodule, and a heavy texture-specific detection submodule. The foreground texture classification and scheduling submodule classifies textures and distinguishes images into light texture images and heavy texture images. The dedicated detection submodule for light textures is adapted to light texture surface scenarios and can detect minute defects. The dedicated heavy texture detection submodule is adapted to heavy texture surface scenes and detects and suppresses false positives in texture detection.

3. The multi-scenario adaptive precision manufacturing high-precision non-destructive defect detection end-to-end system as described in claim 2, characterized in that: Different thresholds are preset for both the light-textured surface scene and the heavy-textured surface scene.

4. The multi-scenario adaptive precision manufacturing high-precision non-destructive defect detection end-to-end system as described in claim 1, characterized in that: The BLI-specific detection model includes a BLI internal foreign object detection submodule and a BRV-GT safety filter submodule. The BLI internal foreign object detection submodule identifies and locates foreign objects in the internal holes. The BRV-GT safety filtering submodule uses a lightweight image processing algorithm to remove the misjudged content from the BLI internal foreign object detection submodule.

5. The multi-scenario adaptive precision manufacturing high-precision non-destructive defect detection end-to-end system as described in claim 4, characterized in that: The BRV-GT security filtering submodule includes the following operating steps: S31. Obtain the dynamic threshold: S32. Noise reduction is achieved by using grayscale binarization and morphological opening operations to segment the effective area of ​​the inner hole from the non-target area of ​​the workpiece and remove small noise points; S33. Check the grayscale distribution of each box and filter out any boxes that fall entirely in low grayscale non-target areas as invalid or incorrect annotations.

6. The multi-scenario adaptive precision manufacturing high-precision non-destructive defect detection end-to-end system as described in claim 1, characterized in that: When the progressive magnification trigger detection mechanism triggers high-magnification detection, a targeted precision detection strategy is adopted; The targeted precision inspection delineates candidate regions based on the spatial distribution of historical defects, and only performs high-magnification precision scanning on the candidate regions.

7. The multi-scenario adaptive precision manufacturing high-precision non-destructive defect detection end-to-end system as described in claim 1, characterized in that: It also outputs the physical dimensions of the defects.

8. A multi-scenario adaptive, high-precision non-destructive testing method for precision manufacturing, characterized in that... It includes the following steps: S1. Acquire workpiece inspection images; S2. Based on the illumination mode, assign the image to the corresponding detection path, i.e., the FLI detection path or the BLI detection path; S3. Analyze the results via the FLI or BLI detection path and output the results; When the result is "No valid defects detected", a high-magnification local fine inspection is triggered and the result is output.

9. The multi-scenario adaptive high-precision non-destructive testing end-to-end method for precision manufacturing as described in claim 8, characterized in that: The FLI detection path includes the following steps: S211. Classify the texture of the image; S212. Based on different texture classification results, assign them to the corresponding detection paths, namely, light texture detection paths or heavy texture detection paths; S213. Analyze the light texture detection path or the heavy texture detection path, and output the results based on the corresponding confidence threshold.

10. The multi-scenario adaptive high-precision non-destructive testing end-to-end method for precision manufacturing as described in claim 8, characterized in that: The BLI detection path includes the following steps: S221. A candidate frame for foreign objects in the internal hole is output by a dedicated detection model; S222. Security verification is performed using the BRV-GT module to filter invalid / erroneous labels.

Images