Identification and detection of pills using deep learning methods.

Dog breed identification is a fine-grained image classification task that requires learning detailed visual patterns across multiple classes. Traditional convolution-based models often face limitations when handling complex variations such as pose, color distribution, and structural similarity among breeds. This project presents a Swin Transformer–based approach for dog breed identification using a shifted window attention mechanism. The proposed method divides images into non-overlapping patches and processes them through hierarchical transformer layers to capture both local and global visual features effectively. The shifted window strategy enables cross-window information exchange while maintaining computational efficiency.

The system is trained using a publicly available dog breed image dataset containing multiple categories with diverse image samples. The model learns discriminative feature representations that improve classification accuracy while reducing overfitting. A Flask-based web application is developed to provide an interactive interface for breed prediction. Users can upload an image, and the system processes it through the trained Swin Transformer model to generate a predicted breed label.

This project demonstrates how transformer-based architectures can be applied to fine-grained visual recognition tasks with improved feature learning and scalability. The results highlight the effectiveness of shifted window transformers in handling complex image classification problems while maintaining modular and deployable system design.

Keywords: Dog Breed Identification, Swin Transformer, Shifted Window Attention, Fine-Grained Classification, Deep Learning, Image Processing, Vision Transformer, Flask Application, Feature Extraction, Neural Networks.