Dog Identification App

Datasets

The Dog Identification App uses two main datasets:

• Dog Image Dataset (Udacity): This dataset consists of images of dogs from 133 different breeds. Download and extract to use in training.
• Preprocessed VGG16 Bottleneck Features: This file contains precomputed bottleneck features using the VGG16 model for faster training.

Results

Below are some examples of predictions made by the Dog Identification App:

Model Explanation

The Dog Identification App uses a convolutional neural network (CNN) with pre-trained VGG16 layers for feature extraction. The extracted features are then passed to a custom classifier that predicts the breed of the dog. Below is a summary of the model:

• VGG16 Feature Extraction: The convolutional layers of the VGG16 model (pre-trained on ImageNet) are used to extract bottleneck features from the input images.
• Custom Classifier: The extracted features are passed to a custom-built classifier with several dense layers and a softmax output, which predicts one of the 133 dog breeds.
• Training Process: The model was trained using the Udacity Dog Images Dataset, with additional preprocessed features to improve efficiency.

Model Comparison

The Dog Identification App was trained using two different neural network architectures: a custom CNN and a pretrained VGG16 model. Below is a comparison of their performance:

Custom CNN Model

The custom CNN model was built from scratch and achieved the following performance:

Though the custom CNN model improved during training, the validation performance showed significant fluctuations, suggesting generalization challenges.

VGG16 Model

The VGG16 model, pre-trained on ImageNet, was fine-tuned for the dog breed classification task. It achieved better performance and more stable validation results:

The VGG16 model outperformed the custom CNN due to its powerful feature extraction capabilities from the pretrained convolutional layers.

Key Observations:

• Accuracy: The VGG16 model reaches a much higher accuracy (~70%) compared to the custom CNN, which peaks at around 20%.
• Loss: The VGG16 model stabilizes the validation loss effectively, avoiding overfitting, while the custom CNN struggles with more erratic loss patterns during validation.
• Generalization: VGG16 demonstrates a higher capacity to generalize across unseen data, making it the superior model in this comparison.