This project classifies nine different types of date fruits using a Custom Convolutional Neural Network (CNN). The custom CNN is lightweight and efficient compared to larger architectures like EfficientNet-B0.
The dataset consists of 9 classes of date fruits:
- Ajwa
- Galaxy
- Mejdool
- Meneifi
- NabtatAli
- Rutab
- Shaishe
- Sokari
- Sugaey
Training Images: 1,156
Testing Images: 502
Images were preprocessed to size 224x224 with augmentation techniques such as:
- Random Horizontal Flip
- Random Rotation
- Normalization
The Custom CNN includes:
- 4 Convolutional Blocks:
- Convolutions → Batch Normalization → ReLU → Max Pooling
- Fully Connected Layer with Dropout (0.5).
Diagram generated using the PlotNeuralNet library
Model Summary:
- Total Parameters: 128,199
- Final Accuracy: 97% on the test dataset.
Model | Parameters | Size (MB) | Accuracy (%) |
---|---|---|---|
Custom Model | 128,199 | 30.22 | 97 |
EfficientNet-B0 | 4,019,077 | 124.56 | 100 |
- Model Size: EfficientNet-B0 is 4.12 times larger than the Custom Model.
- Parameter Count: EfficientNet-B0 has 31.36 times more parameters.