Fruit Ripeness Detection using CNN

Grading fruit by hand is slow, subjective and inconsistent — a packing line needs the same decision on every fruit, every shift. Ripeness is fundamentally a vision problem: color distribution, spotting and texture change predictably as fruit ripens. This project trains a MobileNetV3-Small backbone with two output heads — one for fruit type, one for ripeness stage — on Fruits-360 (131 classes) plus a curated set of about 3,600 banana and mango photos labelled unripe/ripe/overripe by two annotators. Students learn multi-task learning, why color augmentation must be handled carefully when color is the signal, and how a lightweight model is chosen for phone deployment. The demo grades sample fruit photos with representative predictions.

1. Fruits-360 images provide fruit-type labels; banana and mango photos are labelled unripe/ripe/overripe by two annotators.
2. All images are resized to 224×224 and normalized; disagreements between annotators are dropped from the set.
3. A MobileNetV3-Small backbone feeds two classification heads trained with a weighted multi-task loss.
4. Augmentation uses rotation and cropping; color jitter is kept mild because color carries the ripeness signal.
5. Training runs with AdamW and early stopping on validation loss.
6. Both heads are evaluated on a held-out split: accuracy and confusion matrices per head.
7. The web demo grades sample fruit photos through the same pipeline with representative predictions.

Tech stack:

• Python 3, TensorFlow/Keras
• MobileNetV3-Small
• Fruits-360 dataset
• NumPy, Matplotlib, scikit-learn
• Jupyter Notebook (training)
• HTML/CSS/JS grading demo
• TFLite export

Dataset & model details

• Dataset: Fruits-360 (H. Mureșan) — 131 fruit classes for the type head; plus a curated ripeness subset of ~3,600 banana and mango photographs labelled unripe / ripe / overripe by two annotators (disagreements dropped), assembled for the build.
• Task: Multi-task classification; input = 224×224×3 fruit photo; outputs = fruit type (12 classes) + ripeness stage (3 classes).
• Model: MobileNetV3-Small backbone (ImageNet pre-trained) → two heads: Dense(12, softmax) for fruit type, Dense(3, softmax) for ripeness; ~2.5M parameters.
• Metrics: Fruit-type accuracy ~97% (design target), ripeness accuracy ~93% (design target), per-head confusion matrices. No accuracy is claimed as measured until the training run is executed for the order.

• Fruit-type + ripeness classifier Multi-task CNN: one head predicts the fruit, a second head predicts unripe / ripe / overripe — both from a single photo.
• Fruits-360 + ripeness subset 131-class published dataset for fruit identity, plus ~3,600 banana/mango photos labelled by ripeness stage.
• Multi-task training notebook Shared MobileNetV3 backbone, weighted task losses, augmentation, and training curves — fully reproducible.
• Phone-ready model MobileNetV3-Small exported to TFLite (~3.4 MB), chosen deliberately for on-device grading apps.
• Interactive grading demo Single-file web app: pick a sample fruit photo and see the predicted type and ripeness with probability bars.
• Per-stage evaluation Confusion matrices for both heads, so the viva can discuss where ripe/overripe boundaries blur.
• Viva kit Report PDF, PPT and Q&A covering CNNs, multi-task learning, augmentation trade-offs and TFLite deployment.

• Ripeness-labelled banana/mango subset with annotation notes
• Fruits-360 preparation scripts
• Complete multi-task training & evaluation notebook
• Trained model + TFLite export
• Interactive grading web demo with sample photos
• Per-head confusion matrices and training curves
• Project report PDF, PPT presentation
• Viva Q&A preparation document

• Ripeness is graded from appearance only — internal quality (sugar, firmness) needs sensors, not photos.
• The ripe/overripe boundary is genuinely ambiguous; annotator disagreements were dropped, which the report discloses.
• 93% ripeness accuracy is a design target, stated honestly — the report documents the actual figure.
• Unusual lighting shifts predictions; the demo notes the controlled-lighting assumption.
• Only the covered fruit types are supported; new fruits need new labelled data.