Mask constraint-based industrial defect data augmentation system
By using a mask-constrained industrial defect data augmentation system, multi-level spatial constraint masks are constructed to generate defect samples that conform to industrial patterns. This solves the problems of high difficulty and cost in collecting real defect samples and enables efficient multi-task annotation and rapid deployment of AI models.
Patent Information
- Application Number
- CN202611156159.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-10-27
AI Technical Summary
In industrial defect detection, collecting real defect samples is difficult and costly. Traditional data augmentation methods lack industrial spatial constraints, resulting in distorted generated samples and the inability to generate multi-task annotations simultaneously. Consequently, the training data quality of AI detection models is poor and the deployment cycle is long.
An industrial defect data augmentation system based on mask constraints is adopted. A multi-level spatial constraint mask is constructed through a spatial constraint modeling unit. Combined with a defect template management unit, a constraint-based defect generation unit, and a multi-task annotation generation unit, defect samples that conform to industrial patterns are generated and multi-task annotation is performed simultaneously.
It effectively avoids the sample distortion problem caused by the lack of constraints on defect locations in traditional methods, significantly reduces annotation costs and time, improves the adaptability and accuracy of training samples, and shortens the deployment cycle of AI models.
Smart Images
Figure CN122989609A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of industrial machine vision and artificial intelligence training data generation, and particularly relates to a mask-constrained industrial defect data enhancement system. Background Technology
[0002] Industrial defect detection relies on optical imaging and intelligent analysis to automatically identify and determine product appearance defects. It is a core component for ensuring manufacturing yield, reducing inspection costs, and improving the automation level of production lines. As deep learning AI vision inspection gradually replaces traditional manual visual inspection and rule-based solutions, it has become the mainstream technology for detecting small industrial defects. The authenticity, reasonable distribution, and completeness of the training data directly determine the detection accuracy and implementation effect of the model.
[0003] Unlike general natural image vision inspection, industrial defect detection scenarios face multiple inherent industry constraints, constituting the core data bottleneck for AI implementation. First, real defect samples are extremely scarce. While the overall yield of industrial production lines is generally above 99%, the probability of a serious defect is less than 0.1%. Collecting thousands of effective samples requires months and shutdowns, making each sample costly. New production lines lack defect samples before commissioning, hindering timely deployment of AI models. Second, defect distribution is strictly constrained by industrial space, appearing only in specific areas such as solder pads and process gaps. Randomly generated defects do not conform to process rules and lack training value. Third, detection targets are often densely packed micrometer-level targets. A single high-resolution image can contain thousands of detection objects with extremely small adjacent spacing. Defect overlap and boundary violations introduce erroneous features into the model, reducing detection accuracy. Fourth, multi-task annotation is costly. Industrial inspection requires simultaneous coverage of target detection, semantic segmentation, and defect classification tasks. Pixel-level mask annotation is far more expensive than ordinary annotation, and manual annotation has a high error rate. Fifth, systemic defects are difficult to simulate. Industrial production often produces batch defects such as missing balls in rows and regional solder bridging. Traditional methods can only generate isolated single-point defects.
[0004] Current industrial data augmentation techniques are mainly divided into two categories: traditional geometric transformation augmentation and image synthesis augmentation, both of which have insurmountable drawbacks. The former expands data through pixel operations such as rotation, flipping, and adding noise, completely lacking the ability to understand industrial spatial logic. The generated defect locations violate real industrial patterns, easily leading to a high false detection rate in the model. The latter, represented by algorithms such as CutPaste and Copy-Paste, lacks collision detection mechanisms, resulting in easily overlapping and merging defects, and failing to match the defect distribution patterns in real production. This leads to insufficient generalization ability and a significant drop in accuracy after the model is deployed. Furthermore, neither method can synchronously update annotation information; augmented samples still require manual re-annotation, which not only fails to reduce costs but also increases workload and annotation errors. Summary of the Invention
[0005] The purpose of this invention is to provide a Mask-constrained industrial defect data augmentation system to address the technical problems in the field of industrial defect detection, such as the difficulty and high cost of collecting real defect samples, the lack of industrial spatial constraints in traditional data augmentation methods leading to distorted generated samples, the inability to generate multi-task annotations simultaneously, and the resulting poor quality of training data and long deployment cycles for AI detection models.
[0006] To achieve the above objectives, the specific technical solution of the Mask-Constrained Industrial Defect Data Augmentation System of the present invention is as follows:
[0007] A Mask-Constrained Industrial Defect Data Augmentation System includes a spatial constraint modeling unit, a defect template management unit, a constraint-based defect generation unit, an image fusion unit, and a multi-task annotation generation unit.
[0008] The spatial constraint modeling unit is used to construct a multi-level spatial constraint mask adapted to the detection scenario, so as to delineate the allowed and prohibited areas for defect generation in the background image;
[0009] The defect template management unit is used to store and manage industrial defect templates and the associated attribute information of each template;
[0010] The constraint-type defect generation unit is used to generate a defect arrangement scheme within the allowed area using the multi-level spatial constraint mask as the boundary rule.
[0011] The image fusion unit is used to fuse the defect template corresponding to the defect arrangement scheme into the background image and output the defect sample image;
[0012] The multi-task annotation generation unit is used to synchronously generate a multi-task annotation dataset corresponding to the defect sample image based on the defect arrangement scheme and the attribute information of the corresponding defect template.
[0013] As a further improvement of the present invention, the multi-level spatial constraint mask constructed by the spatial constraint modeling unit includes a global constraint mask, a process region mask, and a defect type mask; the global constraint mask is used to mark prohibited areas within the background image range where no defects are allowed to be generated; the process region mask is used to divide process regions with different attributes within the background image; and the defect type mask is used to define the corresponding allowed areas for each type of defect.
[0014] As a further improvement of the present invention, the spatial constraint modeling unit includes a manual mask drawing construction method and an automatic mask generation construction method; the manual mask drawing construction method draws the boundaries of different process areas through a visual interactive tool to generate spatial constraint masks of corresponding levels; the automatic mask generation construction method generates corresponding spatial constraint masks by identifying the positions of functional elements in the scene through a target detection model.
[0015] As a further improvement of the present invention, the defect template management unit adopts a hierarchical and classified storage architecture, including a normal object template library, a defect sample template library and a background template library; each template stores image data, pixel-level segmentation mask, category label and visual feature parameters; the defect template management unit also includes a template enhancement module, which generates corresponding variant templates by performing preprocessing adjustment operations on the original templates to expand the sample diversity of the template library.
[0016] As a further improvement of the present invention, the constraint-based defect generation unit includes a random point generation module, a boundary validity verification module, a defect collision detection module, and an occupancy mask update module; the random point generation module is used to randomly generate the center point of the defect within the allowed area according to a preset defect density distribution rule; the boundary validity verification module is used to verify whether the complete boundary of the defect is within the allowed area of the corresponding category; the defect collision detection module is used to verify the overlap ratio between the current defect and the generated defects, and if the overlap ratio exceeds a preset threshold, it is regenerated; the occupancy mask update module is used to mark the defect area as occupied after the defect is confirmed to be generated, thereby restricting the repeated generation of defects within the area.
[0017] As a further improvement of the present invention, the boundary legality verification module adopts a circumferential sampling verification mechanism: taking the defect center point as the center and the corresponding size of the defect as the radius, a number of boundary verification points are collected on the circumference. When all boundary verification points are within the allowable area, the boundary is determined to be legal.
[0018] As a further improvement of the present invention, the constraint-type defect generation unit further includes a systematic defect generation module, which is used to simulate process defects in industrial production, including defects such as missing rows / columns, regional defects, positional offset defects, and regional size unevenness defects.
[0019] As a further improvement of the present invention, the image fusion unit adopts the Poisson fusion algorithm, which matches the gradient field of the background image and the defect template in the fusion region by solving the Poisson equation, thereby achieving a smooth transition fusion between the defect edge and the background.
[0020] As a further improvement of the present invention, the multi-task annotation dataset generated by the multi-task annotation generation unit includes pixel-level semantic segmentation labels, object detection bounding box labels, defect classification labels, and structured metadata labels; the object detection bounding box labels are compatible with the annotation format of deep learning frameworks.
[0021] As a further improvement of the present invention, the system of the present invention also includes a batch generation control unit, which is used to configure the number of samples generated, defect density and defect type ratio parameters, automatically divide the training set, validation set and test set, generate dataset index file, and support breakpoint resume of generation tasks.
[0022] Beneficial effects:
[0023] By constructing a multi-level spatial constraint mask through a spatial constraint modeling unit, the allowed and prohibited areas for defect generation are clearly defined in the background image. The constraint-based defect generation unit generates a defect arrangement scheme based on the mask as the boundary rule, so that defects are generated only in the legal area. This conforms to the spatial distribution logic of the industrial scenario from the perspective of generation rules, effectively avoiding the problems of unconstrained defect positions and sample distortion in traditional random augmentation, preventing erroneous spatial features from entering the model training stage, and improving the industrial scenario adaptability of the training samples.
[0024] The defect template management unit stores and manages industrial defect templates and their associated attribute information in a unified manner, forming a standardized template data source. This not only ensures the consistency and traceability of generated defect features, but also provides a unified data foundation for subsequent defect arrangement, image fusion, and annotation generation, thereby improving the operational stability and data retrieval efficiency of the entire system.
[0025] The multi-task annotation generation unit can synchronously generate a multi-task annotation dataset that corresponds one-to-one with the defect sample image based on the defect arrangement scheme and template attribute information. This breaks the limitation of traditional data augmentation, which only processes images and cannot update annotations synchronously. It eliminates the need for manual re-annotation after augmentation, significantly reducing the manpower and time costs of annotation. At the same time, it avoids the errors introduced by manual annotation and improves the accuracy and consistency of the annotation data.
[0026] From spatial rule definition, template calling, compliance defect arrangement to image fusion output and annotation generation, a complete automated chain is formed, which can produce labeled industrial defect training samples in batches. This alleviates the pain points of long collection cycles, high costs and insufficient quantities of real defect samples in industrial scenarios. It can provide training data support for AI models in the absence of real defect samples and shorten the deployment cycle of industrial vision inspection systems. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of a Mask-constrained industrial defect data enhancement system according to the present invention. Detailed Implementation
[0028] To better understand the purpose, structure, and function of this invention, a Mask-Constrained Industrial Defect Data Augmentation System of this invention will be described in further detail.
[0029] Implementation example:
[0030] like Figure 1 The system shown is a Mask-constrained industrial defect data enhancement system. In this embodiment, it is applied to the visual inspection scenario of solder ball defects on BGA ball-mounted substrates. It is used to generate labeled training samples for AI detection models. The system includes a spatial constraint modeling unit, a defect template management unit, a constraint defect generation unit, an image fusion unit, a multi-task labeling generation unit, and a batch generation control unit. Each unit is connected sequentially through a data link to collaboratively complete the fully automated generation process from blank substrate input to labeled defect sample output.
[0031] The spatial constraint modeling unit is used to construct a multi-level spatial constraint mask adapted to the solder ball detection scenario, delineating the allowed and prohibited areas for defect generation in the blank substrate background image. In this embodiment, the multi-level spatial constraint mask adopts a three-level mask system, from top to bottom: global constraint mask, process area mask, and defect type mask. The global constraint mask is used to mark the areas within the entire substrate image where no defects are allowed to be generated, specifically including non-process areas such as substrate edges, positioning holes, and mounting device areas. The pixel value of this area is marked as 0, representing that no solder balls or defects can be generated here. The process area mask is used to divide the process function areas with different attributes within the image. In this embodiment, it is divided into three categories: pad area, pad gap area, and blank substrate area, where the pad area is the legal mounting position of the solder ball. The defect type mask is used to define the corresponding allowed generation area for each type of defect. For example, missing ball, off-center ball, and multiple ball defects are only allowed to be generated within the pad area, solder bridging defects are only allowed to be generated in the gap area between adjacent pads, and contamination defects can be distributed within the legal process area of the pads and gaps.
[0032] This unit supports two mask construction methods. The manual drawing method uses a visual interactive tool, allowing operators to draw the boundary contours of different areas on the substrate image. The system automatically fills in and generates the corresponding level of spatial constraint masks, suitable for non-standard and special-specification substrate scenarios. The automatic generation method uses the YOLO object detection model to identify the input blank substrate image, detect the position and contour of all pads, automatically generate pad area masks, and calculate and generate mask data for gap areas and global no-distribution areas based on the pad positions. This method is suitable for standardized mass production line scenarios.
[0033] The defect template management unit adopts a hierarchical and categorized storage architecture to store and manage various templates and corresponding associated attribute information for the ball-laying scenario. Specifically, it is divided into a normal object template library, a defect sample template library, a background template library, and a template enhancement module. In this embodiment, the normal object template library is a solder ball template library, storing real solder ball templates with different diameters, roundness, and reflectivity. The defect sample template library stores various real defect templates such as solder bridging, multiple balls, missing balls, off-center balls, dirt, and solder dross. The background template library stores substrate background templates with different textures and lighting conditions to adapt to different batches and materials of substrates. Each template stores a corresponding defect RGB image, pixel-level segmentation mask, defect category label, and visual feature parameters such as lighting, texture, and edge features. The template enhancement module performs preprocessing operations such as rotation, scaling, brightness adjustment, and contrast adjustment on the original templates to generate a large number of variant templates, significantly expanding the sample diversity of the template library and avoiding feature homogenization in the generated samples.
[0034] The constraint-based defect generation unit uses multi-level spatial constraint masks as boundary rules to generate defect layout schemes that conform to industrial rules within the allowed area. Specifically, it includes a random point generation module, a boundary validity verification module, a defect collision detection module, an occupancy mask update module, and a systematic defect generation module. The random point generation module randomly generates the center point of defects within the allowed area of the corresponding category according to preset defect density distribution rules. In this embodiment, it can be configured to match the actual defect distribution pattern of the ball-mounting head, making the defect density in the substrate edge area higher than in the center area, more closely resembling the real production scenario. The boundary validity verification module verifies whether the complete boundaries of the defects are all within the allowed area of the corresponding category. This module adopts a circumferential sampling verification mechanism: with the defect center point as the center and the corresponding defect size as the radius, 36 boundary verification points are evenly collected on the circumference. When all boundary verification points are within the allowed area of the corresponding category, the boundary is determined to be valid; if any verification point exceeds the allowed area, the center point is regenerated.
[0035] The defect collision detection module is used to verify the overlap ratio between the current defect and all generated defects. When the overlap ratio exceeds a preset threshold of 5%, the position of the current defect is regenerated to avoid overlapping, covering, abnormal adhesion and other forms that do not conform to industrial rules.
[0036] The occupancy mask update module is used to mark the area covered by each defect as occupied in the occupancy mask after each defect is confirmed to be generated, thereby restricting the repeated generation of subsequent defects in this area and further ensuring the non-collision generation effect.
[0037] The systematic defect generation module is used to simulate batch process defects caused by equipment and process deviations in industrial production. Specifically, these include: whole row / column missing defects, simulating whole row / column missing balls caused by pinhole blockage of the ball-placer head; regional defects, simulating regional solder bridging caused by uneven scraper pressure; positional offset defects, simulating whole board solder ball offset caused by conveyor belt offset; and regional size uneven defects, simulating regional solder ball size unevenness caused by uneven distribution of ball-placer head pressure.
[0038] The image fusion unit is used to fuse the defect template corresponding to the defect arrangement scheme into the industrial background image, and output the final defect sample image. In this embodiment, the image fusion unit adopts the Poisson fusion algorithm. The specific execution process is as follows: extract the foreground region of the defect template and the corresponding segmentation mask, calculate the gradient field of the background image in the fusion region, and solve the Poisson equation to achieve a smooth transition between the gradient field of the defect and the gradient field of the background, and finally generate a fused image, so that the lighting, shadow, and edge features of the defect are consistent with the background, eliminating splicing marks and improving the visual realism of the sample.
[0039] The multi-task annotation generation unit is used to synchronously generate a multi-task annotation dataset that corresponds one-to-one with the defect sample images, based on the defect layout scheme and the attribute information of the corresponding defect template. Specifically, it includes four types of annotation data:
[0040] Pixel-level semantic segmentation labels: The segmentation mask of the defect template is mapped to the corresponding position in the sample image to generate a semantic segmentation annotation map with pixel precision.
[0041] Object detection bounding box labels: Calculate the minimum bounding rectangle of each defect based on the segmentation mask, and automatically generate detection label formats compatible with mainstream deep learning frameworks such as YOLO, COCO, and VOC.
[0042] Defect Category Labels: Based on the category attributes of the defect template, each generated defect is automatically assigned a corresponding category label.
[0043] Structured metadata tags: Generate structured JSON files containing information such as defect location coordinates, size parameters, defect type, and confidence level, which facilitates direct integration with AI training systems.
[0044] The batch generation control unit is used to realize the automated batch generation and management of large-scale samples. It supports the configuration of parameters such as the total number of samples generated, the defect density of a single image, and the proportion of each defect type. It can automatically divide the training set, validation set, and test set according to a preset ratio and generate corresponding dataset index files. It also supports breakpoint resume of generation tasks to avoid data loss caused by interruption during the generation process.
[0045] The specific working steps for the 13×15 array BGA ball-mounting substrate in this embodiment are as follows:
[0046] Step 1: Acquire 10 high-resolution images of blank ball-mounted substrates, input them into the spatial constraint modeling unit, and use the YOLO target detection model to identify the position and contour of all pads, automatically generating a three-level spatial constraint mask system consisting of a global constraint mask, a process area mask, and a defect type mask.
[0047] Step 2: Import the original template library containing 400 normal solder ball templates and 20 defect templates into the defect template management unit. Use the template enhancement module to rotate, scale, and adjust the brightness and contrast of the original templates to generate variant templates, thus completing the preprocessing for expanding the template library.
[0048] Step 3: Configure generation parameters through batch generation control unit: Set the total number of images to 10,000, generate 180-210 solder balls per image to match the 13×15 array specification, and set the total defect rate to 5%, of which missing balls account for 2%, off-center balls account for 1%, solder bridging accounts for 1%, and contamination accounts for 1%. At the same time, configure the regional distribution rules of defect density and the dataset division ratio.
[0049] Step 4: Start the generation task. The constraint-based defect generation unit first generates the center point of normal solder balls in each pad area according to the array rules. After the boundary legality is verified and the collision detection is confirmed to be legal, the occupied area is marked by the occupancy mask update module. Then, according to the preset defect ratio, the corresponding number of positions are randomly selected to generate various types of defects. Boundary verification and collision detection are completed one by one during the generation process to ensure that all defects comply with the spatial constraint rules.
[0050] Step 5: Based on the configuration, you can generate systematic defects, such as randomly selecting a column to generate an entire column of missing balls defects, or selecting a local rectangular area to generate a regional solder bridging defect, to simulate systematic process anomalies in real production.
[0051] Step 6: The image fusion unit retrieves the corresponding defect template according to the generated defect arrangement scheme, and uses the Poisson fusion algorithm to fuse all solder balls and defect templates into the blank substrate background image to generate a complete defect sample image.
[0052] Step 7: The multi-task annotation generation unit synchronously generates semantic segmentation labels, target detection bounding box labels, defect classification labels, and structured JSON annotation files for the corresponding samples based on the defect arrangement scheme and template attributes, so as to achieve a one-to-one correspondence between images and annotations.
[0053] Step 8: Repeat steps S4 to S7 to complete the generation of all samples. The batch generation control unit automatically divides the training set, validation set and test set according to the preset ratio, generates the dataset index file, and finally outputs a complete defect dataset that can be directly used for model training.
[0054] The defect samples generated by the system described in this embodiment are 100% located within the legal process area, with no issues such as exceeding boundaries, crossing pads, or abnormal overlap. The defect distribution differs from that in the actual production environment by less than 5%, and most of the generated samples are difficult for humans to distinguish from genuine samples. The AI detection model trained using this generated data significantly outperforms traditional random augmentation schemes in terms of solder ball detection rate and various defect detection rates. The overall false positive and false negative rates are significantly reduced, and the model training cycle can be shortened from the traditional 4-6 weeks to 1-2 weeks. Furthermore, it does not rely on real defect samples and can be quickly deployed in the early stages of new production line commissioning.
[0055] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.
Claims
1. A mask-constrained industrial defect data augmentation system, characterized in that, It includes a spatial constraint modeling unit, a defect template management unit, a constraint-based defect generation unit, an image fusion unit, and a multi-task annotation generation unit; The spatial constraint modeling unit is used to construct a multi-level spatial constraint mask adapted to the detection scenario, so as to delineate the allowed and prohibited areas for defect generation in the background image; The defect template management unit is used to store and manage industrial defect templates and the associated attribute information of each template; The constraint-type defect generation unit is used to generate a defect arrangement scheme within the allowed area using the multi-level spatial constraint mask as the boundary rule. The image fusion unit is used to fuse the defect template corresponding to the defect arrangement scheme into the background image and output the defect sample image; The multi-task annotation generation unit is used to synchronously generate a multi-task annotation dataset corresponding to the defect sample image based on the defect arrangement scheme and the attribute information of the corresponding defect template.
2. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, The multi-level spatial constraint mask constructed by the spatial constraint modeling unit includes a global constraint mask, a process area mask, and a defect type mask; The global constraint mask is used to mark prohibited areas within the background image where no defects are allowed to be generated; The process area mask is used to divide the process areas with different attributes within the background image; The defect type mask is used to define the corresponding allowed area for each type of defect.
3. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, The spatial constraint modeling unit includes a manual mask drawing construction method and an automatic mask generation construction method; The manual mask construction method uses a visual interactive tool to draw the boundaries of different process areas to generate spatial constraint masks of the corresponding level. The automatic mask generation method uses a target detection model to identify the locations of functional elements in the scene and generate corresponding spatial constraint masks.
4. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, The defect template management unit adopts a hierarchical and classified storage architecture, including a normal object template library, a defect sample template library, and a background template library; each template stores image data, pixel-level segmentation mask, category label, and visual feature parameters. The defect template management unit also includes a template enhancement module, which generates corresponding variant templates by performing preprocessing and adjustment operations on the original templates to expand the sample diversity of the template library.
5. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, The constrained defect generation unit includes a random point generation module, a boundary legality verification module, a defect collision detection module, and an occupancy mask update module. The random point generation module is used to randomly generate the center point of the defect within an allowable area according to a preset defect density distribution rule; The boundary legality verification module is used to verify whether the complete boundaries of the defect are all within the allowed area of the corresponding category; The defect collision detection module is used to verify the overlap ratio between the current defect and the generated defects. If the overlap ratio exceeds the preset threshold, the defect is regenerated. The occupancy mask update module is used to mark the defect area as occupied after the defect is confirmed to be generated, thereby restricting the repeated generation of defects in the area.
6. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 5, characterized in that, The boundary legality verification module adopts a circumferential sampling verification mechanism: taking the defect center point as the center and the corresponding defect size as the radius, a number of boundary verification points are collected on the circumference. When all boundary verification points are within the allowable area, the boundary is determined to be legal.
7. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 5, characterized in that, The constraint-based defect generation unit also includes a systematic defect generation module, which is used to simulate process defects in industrial production, including defects such as missing rows / columns, regional defects, positional offset defects, and regional size unevenness defects.
8. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, The image fusion unit employs the Poisson fusion algorithm, which solves the Poisson equation to match the gradient fields of the background image and the defect template in the fusion region, thereby achieving a smooth transition fusion between the defect edge and the background.
9. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, The multi-task annotation dataset generated by the multi-task annotation generation unit includes pixel-level semantic segmentation labels, object detection bounding box labels, defect classification labels, and structured metadata labels; the object detection bounding box labels are compatible with the annotation format of deep learning frameworks.
10. The Mask-Constrained Industrial Defect Data Augmentation System according to claim 1, characterized in that, It also includes a batch generation control unit, which is used to configure the number of samples generated, defect density and defect type ratio parameters, automatically divide the training set, validation set and test set, generate dataset index file, and support breakpoint resume of generation tasks.