Intelligent Defect Generation: Three Steps to a Training Set You Do Not Have

On industrial production lines, high-quality defect data is what trains and optimises AI visual inspection models, and obtaining rare high-quality defect samples takes multi-layered effort. Manually creating defect samples gives low authenticity. High yield rates mean collecting defect samples online is time-consuming. Changing product types and collecting a large number of samples in a short time is challenging, and complex diverse defects make collection inefficient.

KeyeTech's Intelligent Defect Generation Technology, or Defect Synthesis, uses AI and diffusion model technology. Forward and backward diffusion algorithms simulate various types, locations and shapes of defect images from a small number of sample images, and the simulated defects closely resemble actual defects in appearance and characteristics.

It matters at three points in a machine's life. Rapid model construction: with limited rare defect samples, generated defects enable quick model build. Rapid model deployment: when the product type changes, relevant defects are generated for training. Rapid model improvement: when defects are missed on the line, simulated defects are generated to reduce oversights.

The workflow is three steps. Mark defect samples, place good product images, synthesise. Users independently select defect locations, quantities and types. Generated results come with annotation information, so there is no secondary labelling and the images can be used directly for model training.