An implementation of CNN using the natural scenes dataset.
Network Architecture:
The model uses a sequential architecture with Conv2D layers followed by MaxPooling2D layers, a Flatten layer, and Dense layers.
Input Shape: (150, 150, 3)
Output Layer: Softmax activation with 6 units for the 6 classes.
Training Settings:
Optimizer: Adam
Loss Function: Categorical Crossentropy
Metrics: Accuracy
Epochs: 3
Batch Size: 32
Data Augmentation: Enabled during training (ImageDataGenerator with rescaling, shearing, zooming, and horizontal flipping)
2. Results
Accuracy:
- Training Accuracy: Epoch 1: ~63.03% Epoch 2: ~71.95% Epoch 3: ~74.38%
- Validation Accuracy: Epoch 1: ~66.23% Epoch 2: ~72.83% Epoch 3: ~76.40%
- Test Accuracy: 76.40%