Convolutional Neural Network Bird Classifier

Visit the GitHub Repository

Stack

Here are the technologies used in this project:

Overview

• Developed a custom bird classification model using PyTorch with an accuracy of 87% across 20 bird species from a dataset of 3,113 images.
• Focused on optimizing generalization and reducing overfitting, achieving improved performance on unseen test data.

Model Architecture

• Architecture Selection:
  • Experimented with Variational Autoencoder (VAE) and Generative Adversarial Network (GAN).
  • Chose a Convolutional Neural Network (CNN) due to its:
    • Superior feature detection capabilities.
    • Efficiency in training.
    • Proven effectiveness in handling image data.

Sample Code of the Model

class BirdModel(nn.Module):
    def __init__(self, num_classes=20):
        super().__init__()
        # input = 224 x 224 x 3
        self.conv1 = nn.Conv2d(3, 10, 5)
        self.pool1 = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(10, 20, 5)
        self.pool2 = nn.MaxPool2d(2, 2)
        self.fc1 = nn.Linear(20 * 53 * 53, 50)
        self.fc2 = nn.Linear(50, 20)

    def forward(self, x):
        self.to(device)
        x = self.pool1(nn.functional.relu(self.conv1(x)))
        x = self.pool2(nn.functional.relu(self.conv2(x)))
        x = x.reshape(-1, 20 * 53 * 53)
        x = self.fc1(x)
        x = self.fc2(x)
        return x

Training Process

• Data Augmentation Techniques:
  • Implemented:
    • Image rotation.
    • Color fixation changes.
    • Random erasing.
    • Horizontal and vertical flips.
  • Designed to simulate diverse real-world scenarios and improve model robustness.
• Hyperparameter Tuning:
  • Extensive tuning led to a 15% improvement in model precision and recall.
• Loss Functions:
  • Optimized loss functions tailored for classification tasks to enhance performance.

Validation and Overfitting

• Dataset Split:
  • Used a 70-30 split for training and validation.
  • Limitation: Absence of a distinct test set reduces ability to evaluate generalization on unseen data.
• Overfitting Strategies:
  • Implemented data augmentation techniques (e.g., random flips, color jittering).
  • Acknowledged the potential for advanced strategies like dropout or weight decay.

Error Analysis

• Limited error analysis due to the lack of a test set.
• Recognized the importance of a comprehensive error analysis to:
  • Understand misclassifications and biases.
  • Identify patterns in errors.

Future Improvements

• Incorporate a Test Set:
  • Evaluate performance on completely unseen data for better generalization.
• Conduct Detailed Error Analysis:
  • Hypothesize causes of errors and refine the model to enhance robustness.
• Explore Advanced Techniques:
  • Introduce regularization and advanced data augmentation to further mitigate overfitting.