Transformer-Based Models in Image Segmentation and Classification: A New Era in Vision AI

Abstract

Over the past decade, deep learning has revolutionized computer vision, with
convolutional neural networks (CNNs) dominating tasks like image classification and segmentation.
However, a new paradigm emerged as transformer-based models – originally
developed for natural language processing – have begun to surpass previous CNN-based approaches
across vision tasks. This marks a new era in Vision AI, where transformers’ ability
to capture long-range dependencies and global context is reshaping how we design vision
systems. Transformer models have achieved state-of-the-art performance in image classification
(assigning labels to entire images) and segmentation (partitioning images into labeled
regions), often with simpler pipelines and stronger results than their CNN predecessors.