Grading & Sorting

Why Manual Grading Is a Bottleneck

Even in developed markets, a meaningful share of fresh and processed produce is still hand-graded, and operations that automate run into a deeper problem: the trained vision models and NIR calibration curves must be adapted to the specific cultivars, growing conditions, and quality variation of each supply region. The commercial sorting hardware (Tomra, Key Technology, Compac, Unitec) is widely available; the model and calibration layer is where the work is.

A Honeycrisp apple from Washington's Yakima Valley and a Honeycrisp from New York's Lake Ontario region are the same cultivar but can differ significantly in Brix (sugar content), skin color at maturity, and firmness — parameters relevant to grading against USDA grade standards. A vision model trained on one region's fruit misclassifies another region's at a 15-25% error rate. Region- and cultivar-specific training datasets are the foundational investment.

Open data/apple-grading-images.json — it contains metadata and feature vectors from 12,000 Honeycrisp and Gala apple images collected at packing houses in Wenatchee and Geneva, labeled by graders against USDA grade standards: weight class (Extra Fancy/Fancy/No.1), color grade (full color, 50% color, 25% color), external defect flags (bitter pit, bruising, sunburn, lenticel breakdown), and Brix (refractometer measurement on 20% sample).

Computer Vision Architecture for Produce Grading

A production apple grading line vision system has four components working in sequence:

The CNN architecture choice depends on deployment hardware. On a Jetson AGX Orin (common in packing houses due to price point), EfficientNet-B3 at INT8 quantization achieves 45ms inference time vs. ResNet-50 at 78ms, with comparable accuracy.

# Transfer learning setup for apple defect detection
import torchvision.models as models
import torch.nn as nn

backbone = models.efficientnet_b3(pretrained=True)
# Replace classifier head for multi-label defect detection
n_defect_classes = 8  # bitter_pit, bruise, lenticel_breakdown, sunburn,
                       # russeting, insect_damage, dry_skin, color_uneven
backbone.classifier = nn.Sequential(
    nn.Dropout(0.3),
    nn.Linear(backbone.classifier[1].in_features, n_defect_classes),
    nn.Sigmoid()  # Multi-label: each defect independently present or absent
)

# Training data augmentation for packing house conditions
transforms = [
    RandomRotation(360),           # Fruit orientation is random on conveyor
    ColorJitter(brightness=0.3),   # LED illumination variation between packing houses
    RandomHorizontalFlip(),
    GaussianBlur(kernel_size=3),   # Camera vibration from conveyor
    Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
]

Rice and Wheat Grain Quality: Broken, Discoloured, Foreign Matter

Grain quality assessment is the highest-volume grading application — USDA's Federal Grain Inspection Service (FGIS / GIPSA) grades hundreds of millions of bushels of rice, wheat, and corn annually. Traditional visual inspection by human graders is slow (2-5 kg/min), subjective, and vulnerable to fatigue. AI vision systems operate at 100-500 kg/hour with <1% misclassification on trained categories.

Key quality attributes for rice:

USDA FGIS and land-grant universities have published reference standards and labeled datasets for long-grain, medium-grain, and Calrose rice varieties — the best starting point for training before collecting proprietary data.

Prompt: "I have a batch of long-grain rice images captured on a white background conveyor under
diffuse LED lighting [data/rice-grain-images.json — 8,000 labeled grains]. The current model
accuracy is 94% on head rice identification but only 78% on chalky grain detection (our target
is >92% for all classes). Analyze the confusion matrix [attached], identify the failure modes
for chalky grain, and recommend: (1) a data augmentation strategy to address the specific
failure cases, (2) whether a second model with transmitted light images would close the gap,
(3) a threshold calibration approach that minimizes false reject rate (economic cost: $0.22/kg
misclassified as lower grade) while keeping chalky grade-up rate <0.5%."

NIR Spectroscopy: Composition Analysis Without Sampling

Near-Infrared (NIR) spectroscopy measures the absorption of NIR light (780-2500nm) by food samples — different chemical bonds (O-H for water, N-H for protein, C-H for fat/starch) absorb at characteristic wavelengths. The result: a full composition analysis in 30 seconds without sample destruction, at $0.02-0.05/measurement vs. $50-200 for wet chemistry lab analysis.

NIR calibration model development workflow:

1. Collect reference samples (n=200+ for robust calibration)
   → Cover full range of expected composition variation (variety, origin, season)
   → Each sample: NIR spectrum + wet chemistry reference analysis

2. Pre-process spectra
   → Standard Normal Variate (SNV) correction for particle size variation
   → Savitzky-Golay smoothing (2nd derivative) to resolve overlapping peaks
   → Multiplicative Scatter Correction (MSC) for scattering baseline drift

3. Model fitting
   → Partial Least Squares Regression (PLSR): n_components = 6-12 for food matrices
   → Cross-validate: leave-one-sample-out or k-fold (k=10)
   → Report: RMSECV (cross-validation error), R² ≥ 0.95 for quantitative analysis

4. Validation on independent hold-out set (n=50+)
   → RMSEP (prediction error on unseen samples) should be ≤ RMSECV × 1.15
   → Bias check: systematic over/under-prediction indicates subset not in calibration

Open data/nir-spectra-calibration.csv — it contains NIR spectra (1100-2500nm, 2nm resolution) and reference lab values for 350 wheat flour samples from 8 flour mills: moisture%, protein%, ash%, wet gluten%, water absorption (Farinograph). The task: build a robust PLSR model that can replace the Farinograph measurement with a 30-second NIR scan.

NIR applications by commodity:

Automated Sorting Line Calibration and OEE

A sorting line that runs at 90% of rated speed due to miscalibration loses $500-2000/hour in processing capacity (product value × throughput gap). Calibration drift occurs from:

AI-assisted calibration uses reference standard objects (certified color tiles, NIST-traceable size standards) run through the line at shift start. The vision system measures deviation from expected values and adjusts:

Prompt: "Our apple sorting line vision system has been running for 8 months without recalibration.
Calibration tile measurements today [data/calibration-tile-readings.json] show: D65 white tile
luminance down 18% from baseline, red tile chroma shift ΔE=4.2 (CIEDE2000), green tile ΔE=2.8.
Current operational data shows Extra Fancy assignment rate dropped from 42% to 31% over the last
6 weeks (we believe actual incoming quality is stable). Calculate: (1) the correction matrix
needed to restore colorimetric accuracy, (2) whether the Extra Fancy rate drop is fully explained
by illumination drift or indicates a model recalibration is needed, (3) a recalibration schedule
recommendation based on this degradation rate."

Tomato and Grape Grading for Export

Tomato (California processing tomatoes): Export and processing grade requires color uniformity (CIELAB a* value), size (diameter 47-102mm), and absence of blossom end rot, catfacing, and cracking. AI grades 20,000+ tomatoes/hour with <2% misclassification vs. 800/hour by trained human grader at >5% error rate.

Grapes (California table grapes for export): USDA and importer protocols require berry diameter, cluster weight, total soluble solids (Brix ≥16), and defect assessment. NIR on-line Brix measurement eliminates destructive sampling at 100% throughput.

Key Takeaways