计算机科学
人工智能
高光谱成像
卷积(计算机科学)
模式识别(心理学)
特征提取
上下文图像分类
深度学习
卷积神经网络
图像(数学)
纹理(宇宙学)
空间分析
计算机视觉
编码器
语义学(计算机科学)
人工神经网络
深层神经网络
建筑
图像处理
图像纹理
作者
Jing Cheng,Genyun Sun,Aizhu Zhang,Hang Fu,Jiahao Cheng,Zhenxin Shi
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-08-03
卷期号:: 8458-8461
被引量:1
标识
DOI:10.1109/igarss55030.2025.11242728
摘要
Hyperspectral image (HSI) offer extensive possibilities in image classification due to their abundant spectral and spatial details. The U-Net architecture has been shown to be an effective approach for extracting features in HSI classification tasks. However, existing encoder and decoder structures are unable to dynamically capture the distribution of important information in the original data, and they have limitations in exploring both deep and shallow semantic features. To address the limitations of existing methods, we propose a novel dynamic attention U-Net framework for HSI classification, named DA-Unet. This framework emphasizes the extraction and aggregation of both shallow and deep features. In the DA-Unet, we introduce dynamic convolution and spatial-spectral attention mechanisms to dynamically extract and interact with shallow and deep semantic features, thereby capturing spatial details and global features. Experimental results indicate that the DA-Unet demonstrates superior performance in capturing neighborhood detail texture and improving classification accuracy.
科研通智能强力驱动
Strongly Powered by AbleSci AI