|
| 基于注意力机制的低秩自编码器高光谱异常检测 |
| Hyperspectral anomaly detection based on attention mechanism with low-rank auto-encoder |
| 投稿时间：2025-02-25 |
| DOI：10.3969/j.issn.1005-5630.202502250034 |
| 中文关键词: 高光谱图像 异常检测 背景字典训练 自编码器 注意力机制 |
| 英文关键词:hyperspectral image anomaly detection background dictionary training auto-encoder attention mechanism |
| 基金项目: |
|
| 摘要点击次数: 5 |
| 全文下载次数: 7 |
| 中文摘要: |
| 高光谱异常检测作为目标检测领域的重要研究方向，近年来受到广泛关注。尽管基于统计和几何模型的经典算法已取得一定成果，但这些方法难以充分利用高光谱数据中的先验信息，且在捕捉图像空间结构方面存在局限。针对这些问题，提出了一种基于背景字典训练的注意力约束低秩自编码器（attention-constrained low-rank autoencoder，ALAE）用于异常检测。该方法创新性地将低秩先验知识与空间信息相结合：首先，通过高光谱数据分解构建背景字典并训练编码器，充分利用背景的低秩特性；其次，引入注意力机制，有效学习数据的空间结构。实验结果表明，在多个高光谱数据集上，ALAE相较于传统方法及其他基于自编码器的方法展现出更优越的性能，在显著降低误报率的同时提高了检测精度。 |
| 英文摘要: |
| Hyperspectral anomaly detection, as an important research direction in target detection, has gained widespread attention in recent years. Although classical algorithms based on statistical and geometric models have achieved certain results, these methods often struggle to fully utilize prior information in hyperspectral data and have limitations in capturing the spatial structure of the image. To address these issues, an attention-constrained low-rank auto-encoder (ALAE) for anomaly detection is proposed. This method combines low-rank prior knowledge with spatial information. Firstly, a background dictionary was constructed by decomposing hyperspectral data and training an auto-encoder, utilizing the low-rank property of the background. Secondly, the attention mechanism was introduced to effectively learn the spatial structure of the data. Experimental results show that ALAE outperforms traditional methods and other auto-encoder-based methods on multiple hyperspectral datasets, significantly reducing the false positive rate and improving detection accuracy. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
