CropIQ: Fruit & Vegetable Classification

A TensorFlow/Keras project for classifying fruits and vegetables using the Fruits-360 dataset. Trains and evaluates MobileNetV2 and EfficientNetB0 models with proper per-model preprocessing.

Project Structure

Tensorflow/
├── model.ipynb          # Complete training notebook (Google Colab) — 17 code cells + markdown headers
├── README.md
├── .gitignore
├── .gitattributes       # Git LFS tracking for model artifacts
├── models/              # Trained artifacts (Git LFS tracked)
│   ├── final_mobilenet.keras
│   ├── final_efficientnet.keras
│   ├── mobilenet_model.tflite
│   ├── efficientnet_model.tflite
│   ├── labels.txt
│   └── results.json
├── CropIQ/              # Dataset subset (Test/Validation images)
│   ├── README.md        # Fruits-360 dataset documentation
│   ├── Test/
│   └── Validation/

Dataset

Fruits-360 (Version 2026.5.12.0) - 102,551 images across 145 classes (fruits, vegetables, nuts, seeds).

This project uses a 13-class subset + background:

• Apple, Banana, Cabbage, Carrot, Cucumber, Eggplant, Grape, Onion, Orange, Papaya, Pepper, Strawberry, Tomato
• Background (from COCO unlabeled2017)

Split: 50% Training / 25% Validation / 25% Test (preserves original Fruits-360 specimen-level splits)

Models Trained

Model	Preprocessing	Parameters
MobileNetV2	mobilenet_v2.preprocess_input (→ [-1, 1])	~3.5M
EfficientNetB0	efficientnet.preprocess_input (→ [0, 255] internal)	~5.3M

Key Pipeline Fixes Applied

1. Per-model preprocessing (critical): Both models now receive raw [0, 255] images from generators. Each model has a Lambda preprocessing layer:

   • MobileNetV2: scales to [-1, 1] via mobilenet_v2.preprocess_input
   • EfficientNetB0: passes through unchanged; model's internal layers handle scaling

2. Removed rescale=1./255 from all ImageDataGenerator instances — was causing double-scaling for EfficientNetB0

3. BatchNorm freezing during fine-tuning: BN running statistics frozen (layer.trainable = False) on the frozen stem to prevent corruption; BN inside the unfrozen window is trained (batch size 64 keeps stats stable)

4. Original dataset splits preserved: Uses Fruits-360's original Training/Validation/Test folders (specimen-level split by k, k+1, k+2, k+3 rule), not random image-level shuffle

5. Folder→class mapping fixed: Merges Fruit-360 variety folders into their parent classes via prefix matching (e.g. Apple Braeburn → Apple) instead of exact-name matching, which was silently dropping all suffixed varieties (Apple, Grape, Onion, Pepper had zero images)

6. Background replacement as a gated layer: RandomBackgroundReplace is a Layer subclass with a training argument — augmentation runs only during fit(), passes through at inference, and keeps tf.random.uniform out of the exported TFLite graph (was previously breaking converter.convert())

Scalability Optimizations

• Mixed precision (mixed_float16): ~2x faster training on T4 with lower VRAM usage
• Batch size 64: more stable BatchNorm stats, better GPU utilization
• Callbacks on val_accuracy (not val_loss): early stopping + best-weights checkpointing align with the actual goal; TerminateOnNaN guards against divergence
• GPU memory growth + device detection at startup
• Ensemble evaluation: results.json includes averaged-probability ensemble accuracy (typically +1-3% over best single model)

Training Details

Target accuracy: 90-95% on the test set (single models); the ensemble in results.json typically lands at the top of that range.

MobileNetV2

• Phase 1 (Head): 20 epochs, LR 1e-3 → 1e-5 (cosine decay + warmup)
• Phase 2 (Fine-tune last 30 layers): 20 epochs, LR 5e-5 → 1e-6

EfficientNetB0

• Phase 1 (Head): 20 epochs, LR 1e-3 → 1e-5
• Phase 2 (Fine-tune last 30 layers): 25 epochs, LR 5e-5 → 1e-7

Shared: AdamW (weight_decay=1e-4), Label Smoothing (0.1), Class Weights (balanced), Top-3 Accuracy metric, mixed float16

Requirements

pip install tensorflow opencv-python scikit-learn matplotlib tqdm

Usage

Training (Google Colab)

1. Open model.ipynb in Google Colab (GPU runtime)
2. Upload CropIQ.zip to Drive: MyDrive/CropIQ/CropIQ.zip
3. Run cells top-to-bottom (17 code cells with markdown headers) — handles extraction, class merging, background download, training both models, evaluation, ensemble, and export
4. Models saved to Drive: best_mobilenet.keras, best_efficientnet.keras (+ _finetuned variants)
5. TFLite exports: mobilenet_model.tflite, efficientnet_model.tflite, labels.txt
6. Results file: results.json — contains test accuracy, per-class precision/recall/F1, confusion matrices, training history, inference benchmarks, and ensemble accuracy for both models.

Model Artifacts

Trained artifacts are stored in models/ (Git LFS tracked):

# After training, copy from Drive to repo:
cp "/content/drive/MyDrive/CropIQ/*.keras" models/
cp "/content/drive/MyDrive/CropIQ/*.tflite" models/
cp "/content/drive/MyDrive/CropIQ/labels.txt" models/
cp "/content/drive/MyDrive/CropIQ/results.json" models/

Then commit (Git LFS handles large files):

git add models/
git commit -m "Add trained models + artifacts"
git push

Android Integration

For the CropIQ Android app, copy only the TFLite model + labels to app/src/main/assets/:

app/src/main/assets/
├── mobilenet_model.tflite      (or efficientnet_model.tflite)
└── labels.txt

Use the provided CropIQClassifier Kotlin class (supports GPU/NNAPI delegates, Flex ops).

Branch

Active development on cropiq-android-integration branch:

git checkout cropiq-android-integration

Known Limitations

• Background domain shift: Training uses random background augmentation; validation/test use original white studio backgrounds. Real-world deployment (phone camera) will have varied backgrounds.
• External test set needed: 100% accuracy on studio photos ≠ real-world performance. Collect phone photos for true evaluation.
• Eggplant class has 0 samples in current dataset — model cannot predict it. Add Eggplant folders to CropIQ.zip or remove from TARGET_CLASSES.
• TFLite models require Flex delegate (SELECT_TF_OPS) due to mixed-precision training. For pure TFLite, retrain with float32 policy.

License

Dataset: CC BY-SA 4.0 (Mihai Oltean, Fruits-360) Code: MIT License