Unsupervised Defect Segmentation

Unsupervised Defect Segmentation is a model trained on normal data that can identify whether an object is in an abnormal (NG) state, such as damage, deformation, etc.


After completing model training, refer to the video in the Training section to create a dataset version and deploy the model.

Use Cases

Unsupervised Defect Segmentation is suitable for the following scenarios:

  1. Single Object: The dataset contains only one type of object, and its position must remain relatively fixed.

  2. Defect Detection: The object can be in either a normal or abnormal (NG) state, enabling the detection of changes such as damage or deformation.

Model Detection Modes

The new Unsupervised Defect Segmentation model supports four detection modes, which can be selected based on specific requirements:

  1. Image-Level - Whole Image Detection - Analyzes the entire image to determine if defects are present. - Suitable for scenarios where precise localization is not required.

  2. Image-Level - Region Detection - Analyzes specific regions of the image to determine if defects exist within the target areas. - Suitable for scenarios with fixed target regions.

  3. Pixel-Level - Whole Image Detection - Analyzes each pixel in the image and highlights abnormal regions. - Suitable for scenarios requiring precise localization of defect areas.

  4. Pixel-Level - Region Detection - Analyzes each pixel within specific regions of the image and highlights abnormal areas. - Suitable for scenarios requiring precise annotation of target areas.

Annotation Methods

Annotation methods vary depending on the model detection mode, as described below:

  1. Image-Level - Whole Image Detection Annotate each image as either OK (normal) or NG (abnormal).

    ../../_images/img_whole_anno.png

    Supports batch selection of images for unified annotation as OK or NG.

    ../../_images/batch_annotate.png

  2. Image-Level - Region Detection First, define a reference image and select the detection region by drawing a bounding box, then save it as a template. Since unsupervised detection requires objects to remain relatively fixed, the detection regions in all images must cover the same parts.

    ../../_images/img_region_anno.png

    Annotate by clicking on the detection regions and marking them as OK or NG.

    ../../_images/img_region_anno.png

  3. Pixel-Level - Whole Image Detection Use intelligent annotation tools or polygon tools to outline the defect regions.

    ../../_images/pixel_whole_anno.png

  4. Pixel-Level - Region Detection First, define a reference image and select the detection region by drawing a bounding box, then save it as a template.

    ../../_images/pixel_region_anno.png

    Use intelligent annotation tools or polygon tools to outline the defect regions within the detection area. If there are no defects, mark the image as defect-free.

    ../../_images/pixel_region_anno2.png

Notes

  1. Dataset Size Limit The maximum number of images supported for training is 300 images.

  2. Data Consistency Normal images should not include any defect annotations, as this may lead to training failures or suboptimal results.

  3. Multiple Defects Support If an object has multiple defects, you can annotate each defect with a separate region.

  4. Data Ratio - Normal Data Priority: The number of normal images in the training set should be greater than or equal to the number of defect-free images.

  5. Default Configuration The unsupervised defect segmentation model does not apply any data augmentation options by default during training.

  6. Built-in Image Splitting for Training During whole image detection, the training mode can be set to Normal or High Precision. High precision mode splits images into 512×512 or 256×256 resolutions for training, increasing detection time but significantly improving model accuracy.

备注

  • The quality of defect-free data is crucial for unsupervised models; ensure that normal data is accurately annotated.

  • Model performance depends on a reasonable ratio of normal and abnormal images.

Practice

Download the practice data unsupervised_data.zip from the following link: Practice Data

After extraction, you will obtain 11 images and their annotation files (.json). - Upload the images to DaoAI World for annotation practice. - After completing the annotations, upload the images and annotation files together to validate the model training results.