PawPrint-k9-Image-Classifier

• Desktop - https://pawprintk9.streamlit.app/
• Mobile - https://pawprintk9-mobile.streamlit.app/
• Presentation - https://www.canva.com/design/DAGgqwDwixA/aMWKa4rasadbsAA1h_IJDg/edit?ui=eyJIIjp7IkEiOnRydWV9fQ

This repository contains a deep learning model built using Convolutional Neural Networks (CNN) to classify images of dog breeds from the Stanford Dogs Dataset. The model uses the pre-trained MobileNet model for feature extraction and fine-tuning, with custom preprocessing and data augmentation to optimize performance.

Dataset

The Stanford Dogs Dataset (http://vision.stanford.edu/aditya86/ImageNetDogs/) consists of 120 dog breeds from around the world. The dataset contains:

• Number of categories: 120
• Number of images: 20,580
• Annotations: Class labels and bounding boxes

The dataset was sourced from ImageNet and is designed for fine-grained image categorization. Each breed has a subfolder containing images. A dataframe was created by iterating through each image and defining the full image path in a column called image_path. The breed names from the subfolder names were cleaned and assigned to the breed column.

Data Preprocessing

1. Image Resizing & Normalization: The original images varied in size and aspect ratio, so a function was created to resize all images to 224x224 while maintaining aspect ratios. Padding was added when necessary, and pixel values were normalized between 0 and 1 by dividing by 255.

2. Data Split: The dataset was split into 3 subsets:

   • Training: 80% of the images (16,464 samples)
   • Validation: 10% of the images (2,058 samples)
   • Testing: 10% of the images (2,058 samples)

   Additionally, the breed labels were encoded into integers, and a new column, breed_index, was created for model training.

Model Architecture

The model was built using MobileNet—a lightweight, pre-trained deep learning model. MobileNet uses depthwise separable convolutions, which reduce computational complexity without sacrificing performance, making it ideal for this project given the computational limitations.

Image Augmentation

To enhance the model's ability to generalize and deal with a small dataset, image augmentation techniques were applied. This increased the variety of images and allowed the model to learn more robust features.

Transfer Learning Strategy

• Frozen Layers: Initially, all layers of MobileNet were frozen to keep the model lightweight and prevent overfitting. Only the final classification layer was trained.
• Optimizer: The Adam optimizer was used with a learning rate of 1e-4, optimizing for accuracy.
• Callbacks: To prevent overfitting, callbacks such as early_stopping, model_checkpoint, and reduce_lr were used during training.

Training Results (First Model)

• Training time: 1 hour and 31 minutes
• Final Epoch Results:
  • Training Accuracy: 77.15%
  • Validation Accuracy: 80.17%
  • Test Accuracy: 79.5%

Although these results were satisfactory, further optimization was attempted to improve accuracy.

Advanced Model

In an attempt to handle the class imbalance and improve model performance, the following changes were made:

1. Unfreezing the last 20 layers of MobileNet to allow the model to learn from the images directly.
2. Class balancing: The model was adjusted to train more on underrepresented breeds.
3. Training for 50 Epochs.

However, this approach led to some overfitting, as seen by the drop in validation and test accuracy.

Training Results (Advanced Model)

• Training time: 1 hour, 9 minutes, and 27 seconds
• Final Epoch Results:
  • Training Accuracy: 82.09%
  • Validation Accuracy: 73.66%
  • Test Accuracy: 76.58%

Optimized Model

In the final model, the layers were frozen again to balance the dataset, and additional neurons and dense layers were added. This helped to improve the results.

Final Results (Optimized MobileNet Model)

• Training Accuracy: 80.76%
• Validation Accuracy: 80.52%
• Test Accuracy: 80.61%

This result was satisfactory, as the target of 80% accuracy was achieved.

Other Experiments

I also tried training the model using ResNet50, a more complex architecture. However, after 10 epochs, the model showed no learning progress, and I did not have sufficient time to optimize it further.

Model Evaluation

I plotted the evaluation of the best model using:

• Line graphs
• Accuracy reports
• Confusion matrix

Deployment

Finally, the model was deployed on Streamlit Cloud, where it was tested with unseen images. The predictions were accurate, with a confidence level of above 80% for all tested images.

Conclusion

The project demonstrates a successful implementation of a CNN-based image classifier for dog breeds using the Stanford Dogs Dataset. Despite the challenges of class imbalance and limited data, the final model achieved a solid performance and was successfully deployed for real-time predictions.

Requirements

• Python 3.x
• TensorFlow 2.x
• Keras
• NumPy
• Pandas
• Matplotlib
• OpenCV
• Streamlit

Personal Study Only - No Distribution Intended

This notebook contains code and a model developed for my personal study of dog breed classification using the Stanford Dogs Dataset. It is not intended for public use, distribution, or commercial purposes.