AI Harvest Sorting and Quality Assessment System
An optical sorter on grain elevator takes 15 images of each grain in 2 ms and makes rejection decision faster than human blink. This is one of few cases in agro-CV where computer vision has worked in production 20 years—even with classical algorithms. Today's task: raise accuracy, reduce false rejects, predict internal defects.
CV Method Stack by Product Type
Sorting tasks differ radically by product:
Grain and Legumes (High Speed, Unit Processing)
Conveyor optical sorters work at 5–15 t/h. Each grain photographed in RGB + near-IR. Inference must fit in < 3 ms per object. Architecture—light classification network (MobileNetV2 or EfficientNet-Lite0) with INT8 quantization on NVIDIA Jetson or FPGA (Xilinx Zynq).
Wheat grain defect classes: broken, shriveled, sprouted, fusarium-affected, green, crop contaminant. Inter-class confusion—shriveled vs normal under non-standard light. Solution: white reference normalization before each batch + lighting augmentation during training.
Fruits and Vegetables (Appearance + Internal Defect Prediction)
External sorting by size, color, shape—task for classical CV plus detection. More interesting—NIR spectroscopy for non-destructive internal quality: dry matter (°Brix), apple internal browning, hollowness. One NIR sensor (Zeiss Corona 45 NIR or Carl Zeiss MCS 600) + calibration PLS regression—standard on production lines.
Modern approach: hyperspectral camera (400–1700 nm) above conveyor. Each apple receives full spectrum, CNN with 1D-spectral branch predicts °Brix with RMSE=0.6 and internal browning with 0.91 accuracy (vs 0.79 for PLS).
Deep Dive: Potato Sorting
Practical case. Client—potato processor, 8 t/h line. Task: auto-sort into 4 classes (extra / first / second / technical) by combined features.
Features: weight (load cell), size (3D scanner or stereo camera), green areas, eye count and depth, mechanical damage, rot.
Solution architecture:
- YOLOv8m-seg for potato segmentation and defect detection
- Separate regression head for size estimation from mask
- Fuse with weight sensor and color features
- Gradient Boosting for final classification decision across all features
Result: GOST R 51808 standard compliance 94.3% vs 78% prior mechanical screening. False "class downs"—3.1% (important: underestimate costlier than overestimate).
| Metric | Before | After |
|---|---|---|
| Classification Accuracy | 78% | 94.3% |
| Throughput | 8 t/h | 8 t/h |
| False Class Downs | ~12% | 3.1% |
| Manual Control | 2 operators | 0.5 operator |
Integration into Production Line
System works linked with PLC (Siemens S7 / Allen-Bradley) via PROFINET or OPC UA. Rejection air valve command arrives < 5 ms after inference. Conveyor speed—from line controller, CV syncs by encoder.
Deploy: NVIDIA Jetson AGX Orin for heavy models (potato, fruit) or Jetson Orin NX for grain. TensorRT engine, FP16 for accuracy or INT8 for critical speed needs.
Implementation Process
- Line technical survey, identify imaging and lighting points
- Data collection under production lighting, annotation (500+ samples per defect class)
- Train and validate on held-out set
- Integrate with PLC, test at real conveyor speed
- Calibrate rejection thresholds with technologist
- Monitor accuracy in production, retrain on variety/batch change
Timeline
System for one product type (grain or fruit): 6–10 weeks. Complex line with NIR module and ERP integration: 3–5 months.