Case · 2026-05-07

A Fresh-Produce Processor: Integrated Foreign-Object Removal and Quality Grading with Deep Learning

AI-AOI Inspection · Deep-Learning Sorting · High-Throughput Line

Back to Insights

Conventional color sorters let near-colored insects and plastic fragments pass while ejecting blemished but edible produce as defects — detection and false-reject move against each other. This plant rebuilt the decision logic with DaoAI deep-learning models.

>97%异物检出
<1%误剔率
6.5t/h处理量

This fresh-produce processor supplies prepared vegetables and frozen fruit-and-veg ingredients to retail and food-service channels, handling over a hundred tons per day. Conventional photoelectric color sorters rely on fixed color thresholds and are nearly helpless against foreign objects close in color to the raw material — dark insect bodies, translucent plastics, same-colored grit. To avoid misses, operators tightened thresholds, which in turn ejected large volumes of naturally blemished or slightly off-color but perfectly edible produce, keeping good-product loss stubbornly high.

DaoAI deployed an AI-AOI inspection system to replace the rule-based color sorter. The deep-learning model no longer looks at color alone; it combines texture, shape, edge and context to separate true foreign objects from natural raw-material variation — faced with a dark patch, it tells a moldy contaminant apart from a varietal dark spot. For occasional rare contaminant types, APDT few-shot learning lets the line adapt to new raw-material batches from just dozens of samples, with no need to re-annotate massive datasets.

Solution Architecture

  • AI-AOI high-speed line-scan imaging plus deep-learning defect/FO models, interfaced to the existing air-jet ejectors
  • A grading model concurrently outputs size, color and damage grade, so removal and grading complete in one pass
  • APDT few-shot onboarding of new raw-material batches, cutting annotation effort by about 80%
  • All decisions are logged, enabling per-batch traceback of contaminant types and ejection records

Detection and false-reject no longer trade off — the model learned to tell true contaminants from the raw material's natural variation.

After go-live, foreign-object detection rose from roughly 88% on the old color sorter to above 97%, while false-reject fell from 4–6% to under 1% — hundreds of kilograms of good material saved every hour. The line runs stably at 6.5 t/h throughput, with grading and FO removal merged into a single station, manual re-inspection headcount reduced, and overall yield meaningfully recovered.

Book a Demo / Get a Quote View Food & Agriculture solutions