PrimateVision.DeepL

This repository presents custom implementations of the VGG16 and VGG19 convolutional neural network architectures for the classification of monkey species using the 10 Monkey Species dataset.

Dataset

The 10 Monkey Species dataset comprises images of 10 distinct species of monkeys. Each image is tagged with one of the following species:

• 'n0': 'mantled_howler'
• 'n1': 'patas_monkey'
• 'n2': 'bald_uakari'
• 'n3': 'japanese_macaque'
• 'n4': 'pygmy_marmoset'
• 'n5': 'white_headed_capuchin'
• 'n6': 'silvery_marmoset'
• 'n7': 'common_squirrel_monkey'
• 'n8': 'black_headed_night_monkey'
• 'n9': 'nilgiri_langur'

Implementation Details

VGG16 Project

The VGG16 project employs the VGG16 architecture for image classification tasks.

• Optimizer Used: Adam optimizer
• Loss Function: Categorical cross-entropy
• Layers Used: Multiple convolutional layers with 3x3 filters, followed by max-pooling layers to downsample the feature maps. Fully connected layers with dropout regularization are used for classification.
• Regularization Used: L2 regularization with a regularization parameter of 0.01 applied to the dense layers.
• Data Augmentation: Techniques such as random rotations, shifts, flips, and zooms are applied during training to enhance model robustness.

Performance Metrics

• Train Loss: 0.2116 | Train Accuracy: 98.27%
• Validation Loss: 0.4009 | Validation Accuracy: 94.85%
• Test Loss: 0.4911 | Test Accuracy: 91.18%

VGG19 Project

The VGG19 project utilizes the VGG19 architecture for image classification tasks.

• Optimizer Used: Nadam optimizer
• Loss Function: Categorical cross-entropy
• Layers Used: Multiple convolutional layers with 3x3 filters, followed by max-pooling layers to downsample the feature maps. Global average pooling layers are included to reduce spatial dimensions before dense layers with dropout regularization for classification.
• Regularization Used: L2 regularization with a regularization parameter of 0.001 applied to the dense layers.
• Data Augmentation: Similar data augmentation techniques are applied as in the VGG16 project.

Performance Metrics

• Train Loss: 0.3599 | Train Accuracy: 98.72%
• Validation Loss: 0.4815 | Validation Accuracy: 93.38%
• Test Loss: 0.6433 | Test Accuracy: 89.71%

---