Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
A procedurally generated dataset of 15,000 grayscale images depicting industrial metal surfaces with five distinct surface conditions. Purpose-built for training and benchmarking computer vision models in manufacturing quality control applications.
https://img.shields.io/badge/Images-15%2C000-blue" alt="Images">
https://img.shields.io/badge/Classes-5-green" alt="Classes">
https://img.shields.io/badge/Resolution-256%C3%97256-orange" alt="Resolution">
https://img.shields.io/badge/Format-PNG-lightgrey" alt="Format">
https://img.shields.io/badge/Size-%7E390MB-purple" alt="Size">
Defect detection in manufacturing is critical for quality assurance, but real-world defect images are: - ❌ Expensive to collect - ❌ Highly imbalanced (defects are rare events) - ❌ Often proprietary or restricted - ❌ Difficult to reproduce
This synthetic dataset provides a balanced, fully reproducible alternative for developing and testing defect detection algorithms—with no licensing restrictions.
| Class | Description | Train | Val | Total |
|---|---|---|---|---|
| Normal | Defect-free metal surface with natural texture | 2,400 | 600 | 3,000 |
| Scratch | Linear abrasion marks (Bezier curves) | 2,400 | 600 | 3,000 |
| Crack | Branching fracture patterns (L-system) | 2,400 | 600 | 3,000 |
| Rust | Corrosion patches (Perlin noise blobs) | 2,400 | 600 | 3,000 |
| Hole | Puncture defects with depth shading | 2,400 | 600 | 3,000 |
| Total | 12,000 | 3,000 | 15,000 |
✅ Perfectly balanced classes (3,000 images each)
✅ Pre-split into train (80%) and validation (20%)
✅ Rich metadata with 14 features per image
✅ 100% reproducible via seed-based generation (master seed: 42)
✅ CPU-generated using pure NumPy/SciPy—no GAN, no GPU required
✅ Realistic variations in texture type, lighting angle, and noise
├── images/
│ ├── train/ # 12,000 images (80%)
│ │ ├── normal/ # 2,400 images
│ │ ├── scratch/ # 2,400 images
│ │ ├── crack/ # 2,400 images
│ │ ├── rust/ # 2,400 images
│ │ └── hole/ # 2,400 images
│ └── val/ # 3,000 images (20%)
│ ├── normal/ # 600 images
│ ├── scratch/ # 600 images
│ ├── crack/ # 600 images
│ ├── rust/ # 600 images
│ └── hole/ # 600 images
├── metadata.csv # Per-image generation metadata
├── config.json # Full generation configuration
└── README.md # Dataset documentation
| Property | Value |
|---|---|
| Resolution | 256 × 256 pixels |
| Color Mode | Grayscale (L mode, single channel) |
| Format | PNG (lossless compression) |
| Bit Depth | 8-bit (0-255) |
| File Naming | {class}_{index:05d}.png |
Each image has detailed generation metadata in metadata.csv:
| Column | Type | Description |
|---|---|---|
filename | string | Image filename (e.g., scratch_00042.png) |
class | string | Defect class label |
split | string | train or val |
width | int | Image width (256) |
height | int | Image height (256) |
base_intensity | int | Mean grayscale intensity (90-190) |
texture_type | string | Surface texture: brushed, matte, or grain |
lighting_angle | float | Directional lighting angle (0-360°) |
noise_strength | float | Applied noise intensity (0.02-0.20) |
defect_count | int | Number of defects (0 for normal) |
defect_coverage_pct | float | Percentage of image affected by defects |
defect_positions | list | Pixel coordinates of defect centers |
defect_sizes_px | list | Size of each defect in pixels |
generation_seed | int | Seed for exact reproducibility |
generation_timestamp | datetime | When the image was generated |
Train CNNs, Vision Transformers (ViT), or traditional ML classifiers to distinguish between defect types.
Use "normal" images for one-class classifier training; evaluate on defective samples.
Pre-train on this synthetic data, then fine-tune on real manufacturing images.
Compare model architectures with a standardized, balanced dataset.
Study how synthetic data complements limited real-world datasets.
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
transform = transforms.C...
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
## Overview
Weld Surface Defect Detection is a dataset for object detection tasks - it contains Classes annotations for 7,165 images.
## Getting Started
You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.
## License
This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).
Facebook
TwitterThis dataset contains the train/validation split for the NEU Steel Surface Detection Dataset. The original dataset can be found here => https://www.kaggle.com/datasets/rdsunday/neu-urface-defect-database The source also contains the description of the dataset and the type of images it contains.
Accompanying paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8007367/ Accompanying blog post: https://debuggercafe.com/steel-surface-defect-detection/
In short: The dataset contains close up of images of steel surface defects. There are 6 classes into which the defects can be classified. They are:
[
'crazing',
'inclusion',
'patches',
'pitted_surface',
'rolled-in_scale',
'scratches'
]
You can find more details in the paper.
Total samples: 1800 Training samples: 1700 Validation samples: 100
Annotations are in XML format.
Facebook
TwitterStatistics of industrial surface defect detection dataset.
Facebook
Twitterhttp://www.gnu.org/licenses/agpl-3.0.htmlhttp://www.gnu.org/licenses/agpl-3.0.html
This dataset is part of the DAGM 2007 (Deutsche Arbeitsgemeinschaft für Mustererkennung) dataset series, specifically Class10, which focuses on industrial surface defect detection. It is widely used for training and benchmarking machine learning models in computer vision-based quality control systems, including CNNs, attention models, and reinforcement learning agents.
The dataset contains high-resolution grayscale images with and without defects, along with labeled ground truth. Class10 is known for its challenging textures and subtle defect patterns, making it ideal for research on robust and adaptive inspection systems.
This dataset has been cited in various works, including research on weakly supervised learning, attention mechanisms, and real-time defect detection.
Citation: M. Wieler, T. Hahn, and F. A. Hamprecht, “Weakly Supervised Learning for Industrial Optical Inspection,” Heidelberg Collaboratory for Image Processing (HCI), Sep. 14, 2007. DOI: 10.5281/zenodo.12750201
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
## Overview
Surface Defect Detection is a dataset for object detection tasks - it contains Surface Defect Detection annotations for 599 images.
## Getting Started
You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.
## License
This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
## Overview
Metal Surface Defect Detection is a dataset for object detection tasks - it contains Metal Surface Defect Detection annotations for 953 images.
## Getting Started
You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.
## License
This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).
Facebook
TwitterAttribution-NonCommercial-NoDerivs 4.0 (CC BY-NC-ND 4.0)https://creativecommons.org/licenses/by-nc-nd/4.0/
License information was derived automatically
In the development of surface defect detection algorithms for steel, benchmark datasets play a crucial role. However, the widely used benchmark datasets generally suffer from limited scale, insufficient coverage of defect categories, and poor annotation accuracy, making it difficult to meet the growing demand for surface defect detection in steel. To this end, this study constructed and annotated a large-scale, multi-scale benchmark dataset SD10 (Steel Defect Detection) covering 10 typical types of defects: Crazing, Inclusion, Patches, Pitted Surface, Rolled in Scale, Scratches, Blowhole, Break, Fray, and Uneven. The dataset aims to break through the application bottleneck of existing datasets, provide high-quality and reliable data support for in-depth research and optimization iteration of steel surface defect detection algorithms.
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
GC10-DET: Metallic Surface Defect Detection Dataset (Object Detection)
Unofficial redistribution of the GC10-DET metallic surface defect detection dataset, reformatted into a standardized YOLO-compatible directory layout with a deterministic train/val/test split.
Disclaimer
This repository is not an official release of the GC10-DET dataset. GC10-DET was created by Xiaoming Lv, Fajie Duan, Jia-jia Jiang, Xiao Fu, and Lin Gan (Tianjin University), who retain all… See the full description on the dataset page: https://huggingface.co/datasets/dronefreak/GC10-DET.
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
Automated Optical Inspection (AOI) technology plays a significant role in industrial defect detection.However, traditional static optical systems are affected by shadows and surface reflectivity, resulting in a high sensitivity to the direction of the illuminant, false positives, and missed detections, especially for metal parts with complex geometries. Moreover, there is still a lack of large-scale datasets for surface defect detection in such scenarios. To address these issues, an automatic metal surface defect detection technique was proposed based on deep learning and photometric stereo vision, and a Metal Surface Defect Dataset (MSDD) was constructed. Firstly, a Stroboscopic Illuminant Image Acquisition (SIIA) method is proposed, which incorporates a specially designed arrangement of illuminants and a channel mixer to blend the collected multi-channel images into RGB pseudo-color images. Secondly, the MSDD is constructed using this technology. We achieve end-to-end surface defect detection using universal object detectors by mapping color space transformations to spatial domain transformations and employing hue randomization for data augmentation. Finally, four universal object detection methods, including FCOS, YOLOv5, YOLOv8, and RT-DETR are validated on this dataset. The results indicate that these models achieve an average precision of 85.4% on the dataset, significantly outperforming traditional methods.The MSDD consists of a total of 138,585 single-channel images and 9,239 mixed images, including 5746 defect-free images and 3493 images containing a total of eight types of defects. The defect patterns included are generally applicable for the automated visual inspection of casting-formed metal blank surface defects, highlighting its high research value.
Facebook
TwitterAttribution-NonCommercial-NoDerivs 4.0 (CC BY-NC-ND 4.0)https://creativecommons.org/licenses/by-nc-nd/4.0/
License information was derived automatically
In automated surface visual inspection, it is often necessary to capture the inspected part under many different illumination conditions to capture all the defects. To address this issue, at CSEM we have acquired a real-world multi-illumination defect segmentation dataset, called CSEM-MISD and we release it for research purposes to benefit the community.
The dataset consists of three different types of metallic parts -- washers, screws, and gears. Parts were captured in a half-spherical light-dome system that filtered out all the ambient light and successively illuminated it from 108 distinct illumination angles. Each 12 illumination angles share the same elevation level and the relative azimuthal difference between the adjacent illumination angles on the same level is 30 degrees. For more details, please read Sections 3 and 4 of our paper.
The washers dataset features 70 defective parts. The gears and screws datasets feature 35 defective, 35 intact and several hundred unannotated parts. Some defects, such as notches and holes, are visible in most images (illuminations) with intensity and texture variations among them, while others, such as scratches, are only visible in a few.
We split the datasets into train and test sets. The train sets contain 32 samples, and the test set 38 samples. Each sample comprises 108 images (each captured under a different illumination angle), an automatically extracted foreground segmentation mask, and a hand-labeled defect segmentation mask.
This dataset is challenging mainly because:
The dataset is organized as follows:
We provide data loaders implemented in python at the project's repository.
If you find our dataset useful, please cite our paper:
Honzátko, D., Türetken, E., Bigdeli, S. A., Dunbar, L. A., & Fua, P. (2021). Defect segmentation for multi-illumination quality control systems. Machine vision and Applications.
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
## Overview
Steel Surface Defect Detection is a dataset for object detection tasks - it contains Steel Surface Defect Detection annotations for 1,798 images.
## Getting Started
You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.
## License
This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
## Overview
CNC Surface Defect Detection is a dataset for object detection tasks - it contains CNC Product Surface Images annotations for 1,799 images.
## Getting Started
You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.
## License
This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).
Facebook
TwitterApache License, v2.0https://www.apache.org/licenses/LICENSE-2.0
License information was derived automatically
This dataset is a Plastic Gear Surface Defect Dataset, which includes comprehensive defect images of plastic gears from three perspectives: top surface (pin side), bottom surface (gate side), and side surface (tooth surface), with a total of 1165 images and 16 different types of surface defects.
|https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18435255%2Faf48dc744f024e9d744ed9f8a28c2031%2Fsample1.png?generation=1753253986150388&alt=media" alt=""> |
https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18435255%2F834aa4611d8a0a8de7e03d2941d6de3f%2Fsample2.png?generation=1753254001101245&alt=media" alt=""> |
https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18435255%2Fbe68a9f2832cb4274b09a220408bd212%2Fsample3.png?generation=1753254018003695&alt=media" alt="">
Facebook
TwitterThe file contain an experimentally collected dataset of welding defect photos captured with a 50-MP camera for high fidelity. To ensure consistency, samples were placed on a flat surface, the camera was mounted on a tripod 15 cm away and oriented perpendicular to the weld, and focus was set manually on the seam. A certified welding inspector verified all images prior to processing. In total, 71 images were gathered across three defect types: undercut, slag inclusion, and spatter.
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
malleable cast iron surface defect dataset(MCISD)
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
This dataset provides synthetic samples of surface defects generated on a CAD model of a car door. The defects include bumps and peaks, simulated using Free-Form Deformation (FFD) to ensure geometric realism and adaptability to curved surfaces. Surface acquisition is emulated using a virtual 3D profilometric sensor, incorporating both geometric and sensor noise to closely replicate real-world inspection conditions.
All samples are labeled, and the dataset includes depth images, trajectory data, and raw sensor outputs, making it suitable for training and evaluating surface defect detection models in industrial settings.
This dataset is associated with the TriPlay repository on GitHub:
🔗 GitHub Repository
It is also related with the following publication:
📄 Simulation of Laser Profilometer Measurements in the Presence of Speckle Using Perlin Noise
(This dataset is also associated with a manuscript currently under review.)
High-Quality Synthetic Defects: Includes localized surface deformations (bumps and peaks) modeled with Free-Form Deformation.
Virtual Profilometric Scanning: Simulates data acquisition with a 3D profilometer to capture realistic sensor readings.
Realistic Sensor Noise: Adds surface and depth distortion to simulate real acquisition conditions.
Per-Step Trajectory and Sensor Data: Includes detailed trajectory files and raw outputs per scanning step.
Automatically Generated Annotations: Bounding boxes and defect metadata are included for supervised learning.
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
C10-DET represents a dataset designed for the detection of defects on large-scale metallic surfaces. It poses significant challenges in terms of the variety of defect categories, the number of images, and data volume. The presence of surface defects on metallic materials can have detrimental effects on the quality of industrial products, making efficient detection of metallic defects essential to meet the quality standards set by various industries. Consequently, there has been a growing interest in the field of metallic surface defect detection, leading to substantial advancements in quality control for industrial applications. However, the task of identifying metallic surface defects is inherently complex, primarily due to environmental factors like lighting, light reflections, and the unique properties of metal materials. These factors significantly increase the intricacy of surface defect detection. Note, that while the original paper stated that the dataset contained 3,570 grayscale images, the current version offers 2,300 images.
Facebook
Twitterhttps://creativecommons.org/publicdomain/zero/1.0/https://creativecommons.org/publicdomain/zero/1.0/
The dataset is designed to support research and development in strip steel surface defect detection. It contains 2000 records collected from multiple production lines, capturing a variety of operational and environmental conditions. Each record represents a batch of strip steel and includes sensor readings, production metadata, and surface characteristics.
Key attributes include:
Temperature – Operating temperature of the production line.
Pressure – Line pressure during processing.
Speed – Rolling speed of the steel strip.
Vibration – Measured vibration levels in the machinery.
Humidity – Ambient humidity in the production environment.
Previous_Defects – Number of defects detected in prior batches.
Severity_Score – A numerical score indicating the severity of defects.
Production_Line – Identifier for the production line.
Batch_ID – Unique identifier for each batch.
Operator_ID – Identifier for the operator responsible for the batch.
Defect_Type – The target label indicating the type of surface defect. The defect types included are Crazing, Inclusion, Patches, Pitted, Scratches and Rolled.
Additionally, the dataset contains corresponding surface images for each batch, allowing visual inspection of defects alongside operational data. This makes it suitable for approaches that integrate tabular sensor data with image-based observations.
The dataset captures diverse defect representations across multiple production conditions, enabling robust analysis, modeling, and evaluation for quality assurance and surface defect detection in strip steel manufacturing.
Facebook
TwitterDataset from http://faculty.neu.edu.cn/songkechen/zh_CN/zhym/263269/list/index.htm
Kechen Song and Yunhui Yan
Description:
In the Northeastern University (NEU) surface defect database, six kinds of typical surface defects of the hot-rolled steel strip are collected, i.e., rolled-in scale (RS), patches (Pa), crazing (Cr), pitted surface (PS), inclusion (In) and scratches (Sc). The database includes 1,800 grayscale images: 300 samples each of six different kinds of typical surface defects.
Fig. 1 shows the sample images of six kinds of typical surface defects, the original resolution of each image is 200×200 pixels. From Fig. 1, we can clearly observe that the intra-class defects existing large differences in appearance, for instance, the scratches (the last column) may be horizontal scratch, vertical scratch, and slanting scratch, etc. Meanwhile the inter-class defects have similar aspects, e.g., rolled-in scale, crazing, and pitted surface. In addition, due to the influence of the illumination and material changes, the grayscale of the intra-class defect images is varied. In short, the NEU surface defect database includes two difficult challenges, i.e., the intra-class defects existing large differences in appearance while the inter-class defects have similar aspects, the defect images suffer from the influence of illumination and material changes.
http://faculty.neu.edu.cn/_tsf/00/09/FzeYrqRveYRn.jpg" alt="http://faculty.neu.edu.cn/_tsf/00/09/FzeYrqRveYRn.jpg">
For defect detection task, we provided annotations which indicate the class and location of a defect in each image. We have carefully clicked annotations of each target in these images. Fig. 2 shows some examples of detection results on NEU-DET. For each defect, the yellow box is the bounding box indicating its location and the green label is the class score.
http://faculty.neu.edu.cn/_tsf/00/09/ZjUbqmjAJvYz.png" alt="http://faculty.neu.edu.cn/_tsf/00/09/ZjUbqmjAJvYz.png">
Facebook
TwitterAttribution 4.0 (CC BY 4.0)https://creativecommons.org/licenses/by/4.0/
License information was derived automatically
A procedurally generated dataset of 15,000 grayscale images depicting industrial metal surfaces with five distinct surface conditions. Purpose-built for training and benchmarking computer vision models in manufacturing quality control applications.
https://img.shields.io/badge/Images-15%2C000-blue" alt="Images">
https://img.shields.io/badge/Classes-5-green" alt="Classes">
https://img.shields.io/badge/Resolution-256%C3%97256-orange" alt="Resolution">
https://img.shields.io/badge/Format-PNG-lightgrey" alt="Format">
https://img.shields.io/badge/Size-%7E390MB-purple" alt="Size">
Defect detection in manufacturing is critical for quality assurance, but real-world defect images are: - ❌ Expensive to collect - ❌ Highly imbalanced (defects are rare events) - ❌ Often proprietary or restricted - ❌ Difficult to reproduce
This synthetic dataset provides a balanced, fully reproducible alternative for developing and testing defect detection algorithms—with no licensing restrictions.
| Class | Description | Train | Val | Total |
|---|---|---|---|---|
| Normal | Defect-free metal surface with natural texture | 2,400 | 600 | 3,000 |
| Scratch | Linear abrasion marks (Bezier curves) | 2,400 | 600 | 3,000 |
| Crack | Branching fracture patterns (L-system) | 2,400 | 600 | 3,000 |
| Rust | Corrosion patches (Perlin noise blobs) | 2,400 | 600 | 3,000 |
| Hole | Puncture defects with depth shading | 2,400 | 600 | 3,000 |
| Total | 12,000 | 3,000 | 15,000 |
✅ Perfectly balanced classes (3,000 images each)
✅ Pre-split into train (80%) and validation (20%)
✅ Rich metadata with 14 features per image
✅ 100% reproducible via seed-based generation (master seed: 42)
✅ CPU-generated using pure NumPy/SciPy—no GAN, no GPU required
✅ Realistic variations in texture type, lighting angle, and noise
├── images/
│ ├── train/ # 12,000 images (80%)
│ │ ├── normal/ # 2,400 images
│ │ ├── scratch/ # 2,400 images
│ │ ├── crack/ # 2,400 images
│ │ ├── rust/ # 2,400 images
│ │ └── hole/ # 2,400 images
│ └── val/ # 3,000 images (20%)
│ ├── normal/ # 600 images
│ ├── scratch/ # 600 images
│ ├── crack/ # 600 images
│ ├── rust/ # 600 images
│ └── hole/ # 600 images
├── metadata.csv # Per-image generation metadata
├── config.json # Full generation configuration
└── README.md # Dataset documentation
| Property | Value |
|---|---|
| Resolution | 256 × 256 pixels |
| Color Mode | Grayscale (L mode, single channel) |
| Format | PNG (lossless compression) |
| Bit Depth | 8-bit (0-255) |
| File Naming | {class}_{index:05d}.png |
Each image has detailed generation metadata in metadata.csv:
| Column | Type | Description |
|---|---|---|
filename | string | Image filename (e.g., scratch_00042.png) |
class | string | Defect class label |
split | string | train or val |
width | int | Image width (256) |
height | int | Image height (256) |
base_intensity | int | Mean grayscale intensity (90-190) |
texture_type | string | Surface texture: brushed, matte, or grain |
lighting_angle | float | Directional lighting angle (0-360°) |
noise_strength | float | Applied noise intensity (0.02-0.20) |
defect_count | int | Number of defects (0 for normal) |
defect_coverage_pct | float | Percentage of image affected by defects |
defect_positions | list | Pixel coordinates of defect centers |
defect_sizes_px | list | Size of each defect in pixels |
generation_seed | int | Seed for exact reproducibility |
generation_timestamp | datetime | When the image was generated |
Train CNNs, Vision Transformers (ViT), or traditional ML classifiers to distinguish between defect types.
Use "normal" images for one-class classifier training; evaluate on defective samples.
Pre-train on this synthetic data, then fine-tune on real manufacturing images.
Compare model architectures with a standardized, balanced dataset.
Study how synthetic data complements limited real-world datasets.
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
transform = transforms.C...