The nickel-plated steel can of cylindrical cells is both highly reflective and strongly curved — a tough case for metal surface inspection. DaoAI uses multi-light computational imaging to separate illumination from defect signal, stably revealing scratches, dents and stains.
The difficulty of cylindrical-cell appearance inspection concentrates on the nickel-plated steel can: the surface is strongly mirror-reflective, and the cylindrical curvature makes the illumination angle vary continuously with position, so signals from scratches, dents, stains and indentations are often drowned in glare. The battery plant's previous single-light vision scheme nearly failed in the can's specular regions, with misses concentrated on the two sides of the curved surface, while manual re-inspection could not guarantee consistency.
The customer wanted to stably detect appearance defects on the curved metal can without sacrificing cadence, while keeping false judgments within an acceptable range to avoid mass over-rejection of good parts.
DaoAI Solution
DaoAI adopted a multi-light computational imaging scheme: by capturing images under controllable illumination from multiple directions and angles, it exploits the differing response of defects under different lighting to separate surface-defect signals from the specular-reflection background. Combined with an AI-AOI model classifying scratches, dents and stains, and APDT positive-sample learning adapting to surface conditions across can batches, it achieves full circumferential coverage of the curved surface.
- Multi-light, multi-angle controllable illumination capture separates specular reflection from true defect signal
- Computational imaging optimized for cylindrical curvature, covering the full can circumference
- AI-AOI classifies surface defects including scratches, dents, stains and indentations
- APDT positive-sample learning adapts to the reflective state of different can batches
Not making the light brighter, but making the defect appear within the light.
After go-live, judgment accuracy for the cylindrical-cell steel-can surface reached above 98% and defect detection above 99%, with misses on the two sides of the curved surface markedly converging. False judgments were effectively controlled, over-rejection of good parts dropped noticeably, and appearance inspection shifted from manual dependence to stable inline automated judgment.