Title: Analysis Of Images Based On Transformer Architectures
Authors: Askarov Elbek Erkinj?? o'g'li
Volume: 10
Issue: 5
Pages: 497-500
Publication Date: 2026/05/28
Abstract:
This research paper presents a comprehensive examination of the evolution of Transformer architectures, their theoretical underpinnings, and their empirical performance across various datasets - developments that have fundamentally reshaped the landscape of modern computer vision. Although convolutional neural networks (CNNs) dominated visual data processing for over a decade, the rise of Vision Transformer (ViT) models has opened new frontiers in modeling global dependencies and capturing long-range contextual relationships. The study systematically investigates the internal mechanisms of leading approaches - including Vision Transformer, Swin Transformer, DETR, Mask2Former, and the Segment Anything Model (SAM) - with particular focus on self-attention mechanisms, patch embedding strategies, and positional encoding schemes. Comparative statistical evaluations were conducted on widely used benchmarks such as ImageNet, COCO, ADE20K, and Cityscapes, alongside specialized datasets including MedMNIST (medical imaging) and UTKFace (biometric analysis). The findings indicate that while Transformer-based models surpass CNNs in accuracy and mean average precision (mAP) on large-scale datasets, challenges related to overfitting and high computational demands persist on smaller datasets. The paper concludes by outlining promising directions for the practical deployment of Transformers and their growing role in multimodal artificial intelligence systems.