Overview
Date fruit grading is labor-intensive, inconsistent, and slow when done by hand. A Jordanian agricultural producer was manually sorting dates on a production line — with quality varying by worker, shift, and fatigue. Rejected product rates were higher than they should be. Throughput was constrained.
We built a real-time computer vision system that grades dates by ripeness, surface defect, and size as they move down the production line — replacing manual inspection with consistent, high-speed automated sorting.
The Problem
- Manual grading introduced inconsistency: same fruit graded differently by different workers
- Throughput was limited by human inspection speed
- Over-rejection of borderline product was costing revenue
- Client had no visibility into defect patterns over time — no data, no trends
What We Built
Image Capture Setup
- Industrial camera array mounted above the conveyor belt
- Controlled lighting rig to eliminate shadow variation across the production environment
- Calibration pipeline to account for different date varieties (Medjool, Ajwa, Sukkari)
Computer Vision Model
- Custom-trained classification model on a dataset of ~40,000 labeled date images
- Three simultaneous outputs per fruit: ripeness stage (5 classes), defect type (8 classes), size band (S/M/L/XL)
- Real-time inference at conveyor speed: <80ms per fruit
Data Collection & Labeling
- We ran a structured labeling sprint with domain experts from the client's QC team
- Built a lightweight annotation tool for the client's team to continue expanding the dataset
- Synthetic augmentation (rotation, lighting variation, partial occlusion) to improve robustness
Integration & Output
- PLC integration: system sends sort signals directly to the mechanical diverter on the line
- Dashboard: live defect rate, ripeness distribution, and throughput metrics per shift
- Alert system: flags when defect rate spikes above threshold
Results
- Throughput increased by 40% vs. manual inspection
- Consistency: inter-rater agreement improved from ~74% (human) to 97%+ (model)
- Reject rate reduced by 18% through more precise borderline grading
- Defect visibility: client now has daily data on defect patterns they never had before
Notes
Client name withheld under NDA. Deployed in production, Jordan.