Efficiently segmenting images with Vision Transformers: a case study on Swin-UNet

• Ano de defesa: 2025
• Autor(a) principal: Almeida, Pedro Henrique Barbosa de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Atenção com redução espacial; Segmentação; Segmentation; Spatial Reduction Attention; Transformers de visão; Vision Transformers
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-09022026-231446/

Convolution-based models iteratively process local regions of an image, demonstrating a high capacity for extracting short-range spatial features. In contrast, models based on attention mechanisms perform global operations, capable of encoding long-range relationships between input elements. In this context, architectures derived from Vision Transformers (ViTs) have demonstrated superior performance on higher-dimensional images. However, such models have a large number of parameters, which implies higher computational costs in terms of training time and memory usage. This work investigates optimization strategies for training the SwinUNet architecture, based on ViTs, considering the effects of pre-training, the amount of available data, and model dimensionality reduction. Structural modifications to the network were evaluated, as well as the introduction of the Spatial Reduction Attention mechanism, aimed at reducing the number of operations required for inference. The combination of these optimizations resulted in a lighter version of the architecture, characterized by fewer parameters, shorter inference time, and performance comparable to the original model, which, in turn, outperforms strictly convolutional architectures.