Transformer based semantic segmentation model Dong Beom Kim

Abstract

Transformer based semantic segmentation model Dong Beom Kim Dept, of Electrical Engineering The Graduate School Hanyang University In the field of computer vision, semantic segmentation is an important task that allows the extraction of information about objects within an image as well as its contextual information. Conventional models have utilized Convolutional Neural Networks (CNNs) to generate feature maps from images and analyze them using deep neural network models. However, with the advent of Transformer models, which have shown overwhelming performance in the field of natural language processing, there is a growing interest in applying Transformer models across various domains. In the computer vision field, the Vision Transformer, a Transformer-based model, has gained attention for its superior image identification capabilities, surpassing conventional models. Based on this, high-performance models like Segformer and SETR have been proposed for semantic segmentation. However, these Transformer-based models typically have the disadvantage of being heavy. Recently, research has focused on developing lighter models, such as MobileViT and Mobile-former, which maintain the high performance of Transformers while reducing model size. In this paper, we propose the MeTU model, which combines the MobileViT with the U-Net structure, as a lightweight yet high-performance model for semantic segmentation. We conducted a quantitative comparison with other Transformer-based models to validate the superiority of our proposed model. Additionally, using a diverse dataset collected in-house, we demonstrated that our model performs exceptionally well across various image data. As proven by our experiments, the proposed model is lighter than other Transformer-based models and demonstrates high semantic segmentation performance.