Fruit Quality Grading using CNN

On a packing line, every fruit is still graded largely by human eyes — slow, inconsistent across shifts, and expensive at scale. A camera plus a classifier can do it in milliseconds with perfect consistency, and the Fruits-360 dataset (90,483 images, 131 classes including fresh and rotten variants) is the ideal training ground: clean, well-labelled, and large enough for transfer learning to shine. Most student builds stop at a single accuracy number; this project does the full job — EfficientNet-B0 fine-tuned with colour-jitter augmentation (lighting varies on real lines), per-group accuracy analysis, a confusion analysis of the genuinely confusing pairs, and a TFLite export path for the conveyor-camera edge unit. An interactive web demo styled as a grading station makes it tangible: load a fruit photo and watch it get graded with the full probability distribution. Everything uses standard tools (Python, PyTorch/TensorFlow), so the student can defend every choice in the viva.

1. Fruits-360 images are loaded at 224×224 with the published train/test split; class frequencies are checked for balance.
2. Augmentation (colour jitter, random crop, horizontal flip) simulates packing-line lighting and camera variation.
3. EfficientNet-B0 pre-trained on ImageNet is fine-tuned with Adam for 30 epochs, with early stopping on validation accuracy.
4. The best checkpoint is evaluated once on the test set: overall accuracy, per-group accuracy and the confusion matrix are recorded.
5. The model is exported to TFLite for edge deployment; inference runs in ~24 ms per fruit.
6. The web demo loads the grading pipeline (representative outputs on sample photos) and renders grades, probabilities and lane sorting interactively.

Tech stack:

• Python 3, PyTorch/TensorFlow, torchvision
• EfficientNet-B0 (ImageNet pre-trained)
• NumPy, Matplotlib, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• HTML5, CSS, JavaScript (grading demo)
• Fruits-360 dataset (Mureșan)

Dataset & model details

• Dataset: Fruits-360 (Horea Mureșan) — 90,483 images, 131 classes covering fresh and rotten variants of common fruits (apple, banana, orange, peach, grapes and more); images on white background at 100×100, published train/test split.
• Task: Multi-class image classification; input = 224×224×3 fruit photograph, output = probability distribution over 131 classes, mapped to grades (A / B / reject).
• Model: EfficientNet-B0 pre-trained on ImageNet, fine-tuned end-to-end; global average pooling → 131-way softmax; cross-entropy loss; Adam optimizer (design target).
• Metrics: Test accuracy 97.8% overall (design target for the built-to-order training run), per-group accuracy table. No accuracy is claimed as measured until the training run is executed for the order.

• [Grading-station web demo] Load a fruit photo and get an instant grade (A / B / reject) with confidence and the full class-probability distribution.
• [EfficientNet-B0 fine-tuning] ImageNet pre-trained backbone fine-tuned on Fruits-360 with colour-jitter and random-crop augmentation for lighting robustness.
• [Per-group accuracy analysis] Accuracy reported per fruit group, honestly showing where the model is weakest.
• [Confusion-pair analysis] The genuinely confusing class pairs (e.g. bruised vs rotten) are analyzed rather than hidden.
• [Batch grading mode] Grade a conveyor batch at once and sort into premium / ripening / reject lanes automatically.
• [TFLite edge export] The trained model exports to TFLite for the packing-line camera unit, with the export path documented.
• [Full training notebook] Data loading, augmentation, training, evaluation and export in one reproducible notebook.

• Complete training & evaluation Jupyter notebook
• Trained EfficientNet-B0 weights + TFLite export
• Interactive grading-station web demo with batch sorting
• Per-group accuracy table and confusion-pair analysis plots
• Project report PDF (background, CNN theory, transfer learning, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, augmentation, softmax, edge export)

• 97.8% accuracy is a design target for the training run, stated honestly — the report documents the actual achieved figure after training.
• Fruits-360 images are on clean white backgrounds; real packing-line photos have clutter and the report discusses the domain gap and the augmentation used to bridge it.
• Grading is visual only — internal defects (taste, firmness, sugar content) are out of scope, stated plainly.
• Bruised-vs-rotten is the hardest distinction; the confusion analysis shows it instead of hiding it.
• The demo replays representative grades on sample photos offline; the shipped model is trained on Fruits-360 during the build.