Fresh-cut vegetables have a shelf life measured in hours; inspection must be fast, catch millimeter-scale targets, and keep data on-premise. DaoAI moved the model to the line edge.
This prepared-vegetable plant produces salads and ready-to-eat fresh-cut vegetables, with only a few hours between cutting and dispatch, so any line stoppage to wait on cloud inference is unacceptable. Its biggest quality risk is small foreign objects — hairs, tiny insects, torn plastic-packaging fragments — that often occupy only a few dozen pixels in frame. Conventional machine vision and manual inspection miss them easily, and a single complaint hits the brand directly.
DaoAI optimized a lightweight deep-learning model for small-object detection and deployed it on an edge computing box at the line, achieving local real-time inference: images never leave the plant, latency stays in the millisecond range, and the line never slows to wait for a result. Data augmentation and hard-example mining sharpened the model's sensitivity to low-contrast, translucent contaminants; for newly appearing contaminant forms, APDT few-shot learning reinforces the model quickly, avoiding a full retrain every time packaging material changes.
Why It Has To Be at the Edge
- Fresh-cut products have a very short window; cloud round-trip latency would slow the takt, while edge inference enables real-time synchronized ejection
- Image data stays inside the plant, satisfying food-industry data-compliance and confidentiality needs
- The lightweight model runs on low-power hardware, keeping retrofit cost and footprint low
- APDT few-shot brings new contaminant types online within hours
A few-dozen-pixel hair, caught in milliseconds at the edge — speed and precision are no longer a choice of one.
After go-live, detection of millimeter-scale small foreign objects holds steadily above 98%, a marked gain over the prior manual spot-checks and conventional vision, and complaints from missed contaminants dropped sharply. A single line achieves full inspection with no added headcount, and edge deployment makes replicating the setup across multiple lines lightweight and fast.