Cerulean

Cerulean's existing ML model for automated visual inspection of nicotine pouches was unreliable in real production settings. It frequently failed to locate key pouch components, especially when seams were poorly visible or pouches were rotated, making accurate measurement of width, seam widths, and longitudinal seam offset impossible. Processing speed was also inadequate at approximately 12 seconds per image — far too slow for conveyor line operations.

ScienceSoft built a production-ready dual-model computer vision pipeline using YOLO for fast pouch detection and a fine-tuned SAM (Segment Anything Model) for precise segmentation and measurement extraction. The team implemented a reproducible preprocessing pipeline to handle rotation, lighting variation, and partial visibility, and performed iterative training and hyperparameter tuning on 2,000+ production images including low-contrast and edge cases.

The new model achieved 97% accuracy on 2,000 raw conveyor images and was delivered within two months of project kickoff. Image processing speed improved 6x, from approximately 12 seconds to 1.5–2 seconds per image. ScienceSoft also documented maintenance and retraining procedures to help Cerulean sustain model accuracy over time.

• A dual-model architecture (YOLO + SAM) combining detection speed with segmentation precision outperformed a single-model approach for industrial inspection tasks.
• Explicit labeling of seam visibility and iterative training on real-world edge cases (low contrast, rotation, partial occlusion) was key to achieving production-grade accuracy.
• Delivering documentation for ongoing model retraining is critical to long-term reliability in manufacturing QA deployments.