# The Challenge: Microscopic Defects, Massive Consequences But manufacturing them is unforgiving: - **Microscopic scratches or blemishes** alter coating uniformity, introducing variation in sensor calibration. - **Grey-on-black housings** and semi-reflective polymers make contrast low, defeating standard AOI. - **Oil films, static dust, and curved geometries** add unpredictable reflections. Traditional inspection systems—rule-based AOI or even early machine-learning models—struggle to generalize. They need hundreds of images, careful lighting control, and still miss faint low-contrast lines that affect sensor reliability. In EV supply chains where throughput and yield drive cost, this meant rework loops, late defect discovery, and inconsistent quality data across lines. # Why the Problem Is Hard ## Low Optical Contrast Grey coatings on dark housings provide almost no brightness difference; threshold-based detection fails. ## Surface Reflectivity Shiny or semi-matte plastics scatter light unevenly, creating false positives. ## Tiny Defect Scale Scratches can be < 0.1 mm, smaller than a single pixel in older AOI setups. ## Data Scarcity Sensor programs produce many variants but few examples per variant—training a deep model from scratch isn't feasible. # The Overview AI Solution Using the [OV20i Vision System](/content/products/ov20i/index.html), Overview AI engineers deployed a classifier recipe trained with just 8 good and 8 defective samples—a total of 16 images. ## Key Design Principles ### Edge Training All learning occurs locally on the OV20i's NVIDIA Orin NX GPU—no cloud uploads, no latency. ### Smart Illumination Diffuse and coaxial lighting combinations reveal texture differences invisible to the naked eye. ### Adaptive Model Design The classifier learns feature patterns (texture, gloss gradients, edge discontinuities) rather than fixed brightness values, allowing reliable detection across varied lighting. ### Fast Iteration Engineers can relabel or add images directly in a browser interface; each training cycle takes < 30 minutes. **Within a single shift**, the system achieved state-of-the-art accuracy identifying both microscopic blemishes and larger scratches—even on low-contrast grey-on-black surfaces where conventional systems failed. # Results and Impact ### Tiny-Data Efficiency Production-ready model trained from 16 images. ### Lighting Robustness Stable detection under daylight, LED, and overhead fluorescents. ### Consistent Yield Defects caught inline before assembly, preventing downstream module rejects. ### Scalable Deployment Recipe transferable to other sensor or connector variants with minor retraining. For EV suppliers, that means fewer false rejects, faster root-cause feedback, and traceable quality records suitable for PPAP and IATF 16949 audits. # FAQs ### Why are scratches so critical on EV sensor modules? Even microscopic surface damage can alter sealing pressure or interfere with optical/electromagnetic sensitivity, leading to drift or early failure. ### How little data can Overview AI train on? Classifier recipes commonly start from 5–10 images per class. Transfer learning and data augmentation close the gap to large-dataset performance. ### Does the system need controlled lighting? No. The OV-series uses adaptive lighting and exposure control, and the AI learns reflectivity patterns so models remain stable under realistic shop-floor conditions. ### How fast can it be deployed? Typical proof-of-concept to validated model in under 2 hours, including capture, labeling, and training. ### Is this compatible with existing PLC or MES? Yes. EtherNet/IP and PROFINET connectivity allow direct pass/fail output and data logging for traceability.