This study explores the challenge of recognizing American Sign Language (ASL) gestures using deep learning techniques. Gesture recognition remains a significant challenge in assistive technology, where the goal is to develop systems that can accurately and efficiently interpret hand signs. The project aims to address issues such as model generalization, data privacy, and effective classification of complex gesture inputs. To solve this, two deep learning paradigms were compared: Transfer Learning (TL) and Federated Learning (FL). Both models were built using a ResNet50 architecture pre-trained on the ImageNet dataset and fine-tuned to suit the gesture classification task. The TL model was trained centrally using standard supervised learning, while the FL model distributed training across four client datasets, simulating real-world decentralized learning via federated averaging. The system was implemented using TensorFlow and Docker with GPU support and incorporated preprocessing and augmentation techniques to improve generalization. The evaluation compared the performance of the two approaches using standard classification metrics and visual analysis tools. The study contributes a comparative study between centralized and decentralized learning methods in gesture recognition, demonstrating their potential and trade-offs. Key findings include a 99.69% validation accuracy for the TL model on subset 2, highlighting its effectiveness in centralized settings. The FL model, while starting at a lower accuracy, improved over multiple rounds—reaching up to 88.46% accuracy, in subset 2, after eight rounds of aggregation—demonstrating the promise of privacy-preserving, distributed learning strategies.