Fruit Freshness Detection using Classical Computer Vision & SVM

A classical computer vision + machine learning project that classifies fruit type (apple, banana, orange) and predicts freshness (fresh / rotten) using a handcrafted feature pipeline and Support Vector Machines (SVM).

This project intentionally avoids deep learning to emphasize interpretability, data discipline, and reproducibility.

---

🔗 Dataset (Kaggle)

Fruit Freshness Dataset (Apple, Banana, Orange) 👉 https://www.kaggle.com/datasets/user2036/fruit-freshness-dataset-v1

Current dataset version: v2

• Removed exact duplicates (file hash)
• Removed perceptual duplicates (pHash)
• Verified no train–test leakage
• Cleaned, frozen, and versioned dataset

All experiments and results in this repository are based strictly on Kaggle v2.

---

Key Highlights

• Fully CPU-friendly (no GPU, no CNNs)

• 30-dimensional interpretable handcrafted feature vector

• Two independent SVM classifiers:

  • Fruit classifier → apple / banana / orange
  • Freshness classifier → fresh / rotten

• Explicit dataset sanitation pipeline

• Stratified K-Fold cross-validation

• Learning curve analysis

• CLI-based training, evaluation, and prediction

---

Final Project Structure (Accurate)

mini-project/
├── clean_dataset/
│   ├── find_image_duplicates.py
│   ├── keep_best_train_duplicates.py
│   └── move_test_leaks.py
│
├── dataset/
│   ├── train/
│   │   ├── apple/
│   │   ├── banana/
│   │   ├── orange/
│   │   ├── rottenapples/
│   │   ├── rottenbanana/
│   │   └── rottenoranges/
│   ├── test/
│   │   └── (same structure as train)
│   └── dataset-metadata.json
│
├── duplicates/
│   ├── leak_groups/
│   ├── train_groups/
│   ├── post_clean_check/
│   └── duplicate_report.csv
│
├── models/
│   ├── fruit_type_svm.joblib
│   ├── freshness_svm.joblib
│   ├── label_encoder.joblib
│   ├── feature_schema.json
│   ├── confusion_matrix_fruit.png
│   ├── confusion_matrix_freshness.png
│   └── learning_curve.png
│
├── src/
│   ├── extract_features.py
│   ├── train_svm.py
│   ├── evaluate.py
│   ├── cross_validation.py
│   ├── learning_curve.py
│   ├── predict_cli.py
│   ├── save_feature_schema.py
│   └── utils.py
│
├── dataset_snapshot.txt
├── requirements.txt
└── README.md

---

Dataset Freezing & Integrity

• dataset_snapshot.txt records the exact directory structure

• Duplicate detection performed using:

  • File hash → exact duplicates
  • Perceptual hash (pHash) → near-duplicates

• Post-clean verification confirms:

  • ✅ No remaining duplicates
  • ✅ No train–test leakage

• Dataset is treated as read-only after freezing

---

Feature Extraction (Frozen Schema)

The system uses a fixed 30-feature vector, stored in:

models/feature_schema.json

Feature categories

Color

• RGB mean & standard deviation
• HSV circular mean & std
• LAB mean & std

Texture

• Laplacian variance
• GLCM contrast, energy, homogeneity
• Grayscale entropy

Shape

• Area, perimeter
• Circularity, solidity
• Aspect ratio, extent

Decay

• Dark pixel ratio

⚠️ Feature order and definitions must not be changed without retraining.

---

Segmentation Strategy

1. Convert to grayscale
2. Gaussian blur
3. Otsu thresholding
4. Largest contour selected as fruit mask

If segmentation fails, the pipeline falls back to whole-image statistics to avoid crashes.

---

Training

python src/train_svm.py

Training set size: 9,453 images

What happens:

• Feature extraction from dataset/train
• Label encoding for fruit classes
• GridSearchCV with RBF-kernel SVM
• Best models saved to models/

---

Evaluation

python src/evaluate.py

Test set size: 1,879 images

Fruit classification

• Accuracy: 97.98%
• Macro F1: 0.98

Freshness classification

• Accuracy: 96.17%
• Macro F1: 0.96

Confusion matrices are saved to:

models/confusion_matrix_fruit.png
models/confusion_matrix_freshness.png

---

Cross-Validation

python src/cross_validation.py

• Stratified 5-Fold CV on training data
• Mean accuracy: 97.45%
• Standard deviation: 0.0036

This confirms model stability and absence of data leakage.

---

Learning Curve

python src/learning_curve.py

Generates:

Learning Curve (SVM)

Observations:

• Validation accuracy plateaus after ~60% data
• Removing ~2,000 duplicates did not reduce performance
• Confirms dataset is information-sufficient, not overfitted

---

Prediction (CLI)

python src/predict_cli.py --image /absolute/path/to/image.jpg

Example outputs:

Fruit: banana (88.80%)
Freshness score: 68.79% → FRESH

Fruit: apple (99.88%)
Freshness score: 0.01% → ROTTEN

The CLI reports class probabilities, not just labels.

---

Performance Summary (Clean Dataset)

Task	Accuracy
Fruit classification	~98%
Freshness classification	~96%

Results remain stable after duplicate removal, validating dataset quality.

---

Known Limitations

• White-background bias in dataset
• Sensitivity to lighting and color casts
• Early decay can be visually subtle
• Classical CV has a bounded ceiling vs deep learning

These are expected, documented trade-offs, not bugs.

---

Why Classical CV + SVM?

This project prioritizes:

• Interpretability over raw accuracy
• Data discipline over brute force
• Reproducibility over black-box models

It is designed as a strong baseline and an educational reference.

---