Fruit Freshness Classification using CNN

Fruit grading in retail shops, mandis and small processing units is still done by eye: a worker glances at each batch and decides what stays on the shelf. That judgment is inconsistent — it varies with lighting, fatigue and experience — and mistakes cost money in both directions, with rotten fruit reaching customers or good fruit thrown away. Computer vision offers a repeatable alternative, and freshness classification is an ideal teaching problem because the visual cues (discoloration, spots, mold, shriveling) are exactly what convolutional networks learn well. This project builds that classifier properly: a MobileNetV2 backbone fine-tuned on the public Fruits Fresh and Rotten dataset (six classes across apples, bananas and oranges), with documented preprocessing, augmentation, training and evaluation. A web demo makes the model tangible — select a fruit photo and watch the network return its verdict with confidence scores.

1. The Fruits Fresh and Rotten dataset (~13,600 images, 6 classes) is loaded and split into training (~10,900) and held-out test (~2,700) sets, with a validation split carved from training.
2. Every image is resized to 224×224×3 and normalized; augmentation (random flips, ±20° rotations, brightness/zoom jitter) is applied during training.
3. A MobileNetV2 backbone pretrained on ImageNet is attached to a new head (GlobalAveragePooling → Dropout → Dense(128) → Softmax(6)); the head trains first, then the top backbone blocks are unfrozen for fine-tuning.
4. Training uses Adam with a low learning rate and early stopping on validation loss; the best checkpoint is kept.
5. The held-out test set is evaluated once: overall accuracy, confusion matrix and per-class precision/recall/F1 are generated for the report.
6. In the web demo, a selected photo passes through the identical preprocessing and the saved model, and the app renders the predicted class with its full probability distribution and a fresh/rotten verdict.

Tech stack:

• Python 3, TensorFlow/Keras
• MobileNetV2 (transfer learning)
• NumPy, OpenCV (preprocessing)
• Matplotlib, scikit-learn (evaluation)
• Jupyter Notebook (training)
• HTML/CSS/JavaScript (web demo)
• Fruits Fresh and Rotten dataset (Kaggle)

Dataset & model details

• Dataset: Fruits Fresh and Rotten for Classification — Kaggle (Sriram R.). Approximately 13,600 images in 6 classes: freshapples, freshbanana, freshoranges, rottenapples, rottenbanana, rottenoranges; pre-split into ~10,900 train / ~2,700 test.
• Task: 6-class image classification; input = 224×224×3 RGB image, output = probability distribution over the 6 classes, mapped to a fresh/rotten verdict.
• Model: MobileNetV2 backbone (ImageNet pretrained) → GlobalAveragePooling → Dropout(0.3) → Dense(128, ReLU) → Dense(6, softmax); two-phase training (head first, then fine-tune top blocks).
• Metrics: Test accuracy (design target ≈ 97%), per-class precision/recall/F1, confusion matrix. No accuracy is claimed as measured until the training run is executed for the order.

• [6-class CNN classifier] MobileNetV2 backbone (ImageNet weights) with a custom classification head — distinguishes fresh vs rotten for apples, bananas and oranges in one model.
• [Live web demo] Select any of the built-in sample photos (or upload your own) and get the predicted class, freshness verdict and full per-class probability bars.
• [Full training notebook] Dataset loading, 224×224 preprocessing, augmentation (flips, rotations, brightness shifts), transfer-learning setup, fine-tuning schedule and evaluation in one reproducible Jupyter notebook.
• [Confusion matrix & per-class report] Precision, recall and F1 for all six classes, so the viva can discuss exactly which classes confuse the model (typically fresh vs rotten oranges at the boundary).
• [Training curves] Accuracy and loss plots for train vs validation across epochs, with interpretation notes in the report.
• [Augmentation pipeline] Documented augmentation policy that teaches the model to tolerate varied lighting and backgrounds instead of memorizing the training set.
• [Exported trained model] Saved weights plus the preprocessing pipeline, so the demo runs the real network without retraining.

• Complete training & evaluation Jupyter notebook
• Trained MobileNetV2-based model file with preprocessing code
• Web demo wired to the trained model (sample photos + upload)
• Confusion matrix, per-class metrics and training-curve plots
• Project report PDF (background, CNN theory, architecture rationale, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, augmentation, softmax, overfitting)