Transformer-Based Deep Learning Architectures for Computer Vision: A Comprehensive Survey and Analysis
Main Article Content
Abstract
The advent of transformer architectures has fun-damentally reshaped the landscape of deep learning, originally demonstrating remarkable success in natural language processing tasks before making substantial inroads into computer vision. This paper presents a comprehensive survey and analysis of transformer-based deep learning architectures and their applica-tions within computer vision domains. We systematically review key architectural innovations, including the Vision Transformer (ViT), Swin Transformer, and their numerous derivatives, exam-ining their design principles, strengths, and limitations compared to traditional convolutional neural network (CNN) approaches. Furthermore, we discuss the integration of self-attention mechanisms into object detection, semantic segmentation, and image generation pipelines. We analyze performance benchmarks across widely adopted datasets such as ImageNet, COCO, and ADE20K, drawing comparative evaluations between transformer-based and CNN-based models. Additionally, we explore training strategies, data augmentation techniques, and transfer learning paradigms that underpin the success of these architectures. Finally, we highlight open research challenges and identify promising fu-ture directions, including efficient attention mechanisms, hybrid architectures, and selfsupervised pre-training. This survey aims to serve as a structured reference for researchers and practi-tioners seeking a rigorous understanding of the current state of transformer-based deep learning for computer vision.