Computer vision · Deep learning · 2022
Wise Lemon Selector
Computer-vision and clustering application that grades the quality of lemons from images.
- 98%accuracy on the test set (380 images)
- 3CNN-based systems: from scratch and fine-tuned (VGG16)
Highlights
- Built three CNN-based systems reaching over 95% precision, both from scratch and with transfer learning and fine-tuning on well-known networks such as VGG16.
- Used batch normalisation, dropout and image preprocessing to prevent overfitting.
Results
| Class | Precision | Recall | F1-score | Images |
|---|---|---|---|---|
| Bad quality | 0.99 | 0.96 | 0.97 | 159 |
| Empty background | 0.99 | 1.00 | 0.99 | 69 |
| Good quality | 0.97 | 0.99 | 0.98 | 152 |
| Accuracy | 0.98 | 380 | ||
| Macro average | 0.98 | 0.98 | 0.98 | 380 |
| Weighted average | 0.98 | 0.98 | 0.98 | 380 |
Technologies
- Computer vision
- CNN
- Transfer learning
- VGG16
- TensorFlow
- Clustering