Only three photos of a rare crack? Synthesize cracks on good-part images with IntSynthesis to train, and keep the real cracks apart to validate on.
EP.09 · Getting started · Only 3 Crack Photos? Defect Synthesis for AI Inspection
Micro-cracks in semiconductor IC packages: a defect so rare there are only three crack photos to train an AI on.
Rare defects leave too little training data. Synthesize cracks on good-part images to train with, and keep the real photos for validation. Three is the film's example, not a required number or a performance promise.
Synthetic cracks for training, real ones for validation.
Chapters: 0:00 Rare cracks · Not enough AI training data? / 0:16 Defect synthesis · IntSynthesis and augmentation / 0:27 Optics Lab · Compare lighting, cameras and previews / 0:37 Train with synthetic data, validate on real cracks
Can I apply the fix from a video to my line as is? Each video is a short animation that walks through a problem and a fix. Whether it works on your line depends on a review of samples, imaging conditions and the process.
Every Tuesday 11:00 KST (02:00 UTC), a new episode premieres on YouTube.
* Every episode is an animated dramatization. CEO Na's voice is an AI clone of the CEO's own, used with consent; other voices are AI-synthesized and the music is AI-generated. On-screen values, charts and digital-twin scenes are examples; defect photos and heatmaps are inspection examples. IntLearn ML autonomous control is under development and needs on-site validation.