wings bird classifier
Image classification pipeline utilizing custom CNNs and transfer learning (ResNet-50) to identify bird species.
December 2025 - Present
Wings is a deep learning project focused on bird species classification. The project explores the performance trade-offs between building a custom Convolutional Neural Network (CNN) from scratch and leveraging industry-standard transfer learning models like ResNet-50.
Technical Insights
Designed and implemented an end-to-end image classification pipeline using PyTorch and torchvision. The project involved comparing a ground-up architecture with a transfer learning approach to evaluate training efficiency and generalization.
- Custom CNN Development: Implemented a modular architecture using convolution, batch normalization, ReLU activation, max pooling, and fully connected layers.
- Transfer Learning: Applied a pretrained ResNet-50 backbone, replacing the classification head for specific bird species prediction.
- GPU Acceleration: Enabled seamless switching between CPU and CUDA for accelerated training and inference.
- Optimization Techniques: Mitigated overfitting through dropout, data normalization, and controlled freezing/unfreezing of model layers.
Results
The study achieved faster convergence and improved accuracy using pretrained models compared to the custom architecture on limited datasets. This project highlights a scalable system design that allows for extension to new datasets or deeper network architectures.
