Associate professor
Supervisor of Master's Candidates
Hits:
Affiliation of Author(s):计算机科学与工程学院
Journal:2023 4th International Conference on Computer Vision, Image and Deep Learning
Funded by:自选课题
Key Words:remote sensing; semantic segmentation; channel; Transformer; CNN
Abstract:Remote sensing semantic segmentation is applied to land cover mapping and urban change detection. Recently, some methods introduced CNN and Vit hybrid methods to solve the remote sensing semantic segmentation. Furthermore, to reduce the computation cost, they mainly introduce the window-self-attention or not the villain self-attention to extract the global semantic feature. However, the window-self-attention ignores the channel interaction information. To fix the problem, we propose Channel-wise Unet-based Transformer (CUformer). The CUformer introduces channel attention to add the channel interaction information into the window-self-attention. Furthermore, previous works have proved that semantic segmentation needs to fine-grain the global semantic feature. Therefore, we design a simple fine-grained head, which introduces a local feature to fine grain the global semantic feature. Extensive experiments show that our model achieves state-of-the-art performance in Potsdam and Vaihingen datasets.
Co-author:胡统业,鲍学良,仲昭昊
First Author:Kevin Liu
Indexed by:Essay collection
Page Number:2
Translation or Not:no
Date of Publication:2023-05-14
