AI Vision for Foam Mold Clip Detection
Achieve 100% accuracy detecting clips in automotive & aerospace molds despite resin, vibration, and variation.
TL;DR (Quick Answer)
A Tier-1 supplier's legacy vision system failed to detect black clips in foam molds due to resin buildup and vibration. Overview.ai’s OV20i Vision System, using a classifier recipe, achieved 100% accuracy with just a dozen training images, providing a stable solution without needing recalibration.
Situation
Detecting black clips embedded in foam molds may sound like a simple presence/absence task. But in production environments, variation breaks systems long before accuracy does. This is a common challenge in both automotive and aerospace assembly where component verification is critical.
The Problem
A Tier-1 automotive and aerospace supplier faced this issue with eight different mold variations. Their $100K legacy system constantly flagged false rejects and couldn’t be trusted for unattended operation due to several factors:
- Resin buildup altered the mold’s appearance over time.
- Reflections and lighting changed with every cycle.
- Vibration from nearby presses created instability and inconsistent imaging.
- Mold-to-mold variation made rule-based systems brittle.
The Overview.ai Solution
Using the OV20i industrial vision system, engineers trained a simple classifier recipe. Within hours, the model reached 100% accuracy on presence/absence checks and learned to classify *mis-seated clips* separately.
Fast Training with Classifier Recipes
With just a dozen labeled images, the on-camera AI learned to distinguish between "clip present," "clip absent," and "clip mis-seated," demonstrating the power of small, high-quality datasets.
Robust Optics & In-Camera Processing
Key to success was robust optics and a segmentation-grade sensor that normalized for lighting and surface texture changes. The system stayed stable through resin buildup and vibration—no recalibration required.
Real-time Feedback with Node-RED
The inspection results were integrated with the line controller using Node-RED logic, providing immediate pass/fail feedback and closing the loop for operators.
Key Engineering Takeaways
- Small, high-quality datasets outperform large, noisy ones when lighting is consistent.
- Stability against vibration and surface contamination is as critical as accuracy.
- Real-time feedback through Node-RED logic closes the automation loop for operators.
FAQ
How many samples are typically needed to train for new molds?
10–15 representative samples are usually sufficient if lighting and optical parameters remain stable.
Can this handle resin discoloration or partial fill?
Yes — segmentation features in the OV vision systems can adapt to grayscale shifts and texture changes with minimal retraining.
Ready to Automate Your Clip & Component Detection?
Get started with the OV20i AI Vision System for reliable, high-speed assembly verification.