摘要
现阶段Transformer模型的应用提升了高光谱图像去噪的性能,但原始Transformer模型对图像空间-光谱耦合关联性的利用仍存在不足;对空间特征的处理存在过于平滑,容易丢失小尺度结构的现象;同时在光谱维度上也过于关注全部通道特征,缺乏对不同光谱波段间差异性的利用;为了应对这些问题,本文提出了一种新的稀疏空谱Transformer模型,提升了对空谱耦合关联性的利用。在空间维度,引入局部增强模块增强空间特征细节,应对过平滑问题;同时在光谱维度上提出了Top-k稀疏自注意力机制,自适应选择前K个最相关的光谱通道特征进行特征交互,从而能够有效捕获空谱特征。最终通过稀疏空谱Transformer的层级残差连接实现高光谱图像的去噪。在ICVL数据集上分别对高斯噪声和复杂噪声进行去噪处理,峰值信噪比分别达到40.56 dB和40.19 dB,证明了本文提出的稀疏空谱Transformer模型优越的性能。
The application of Transformer models has improved the performance of hyperspectral image denoising.However,the original Transformer model still falls short in effectively leveraging the spatial-spectral coupling in HSIs.It tends to excessively smooth spatial features,leading to the loss of small-scale structures.Moreover,it overly emphasizes all spectral channel features,neglecting the differences between different spectral bands.In order to solve these problems,this paper introduces a novel Sparse Spatial-Spectral Transformer model,enhancing the utilization of spatial-spectral coupling.In the spatial dimension,a local enhancement module is introduced to refine spatial feature details and deal with over-smoothing problem.Simultaneously,in the spectral dimension,a Top-k sparse self-attention mechanism is proposed,which adaptively selects the top-K most relevant spectral channel features for feature interaction,effectively capturing spatial-spectral characteristics.Ultimately,hyperspectral image denoising is achieved through hierarchical residual connections with the Sparse Spatial-Spectral Transformer.On the ICVL dataset,denoising performance for both Gaussian noise and complex noise attains peak signal-to-noise ratios of 40.56 dB and 40.19 dB,respectively,demonstrating the superior performance of the proposed Sparse Spatial-Spectral Transformer model in this paper.
作者
杨智翔
孙玉宝
白志远
栾鸿康
Yang Zhixiang;Sun Yubao;Bai Zhiyuan;Luan Hongkang(School of Computer Science&School of Cyberspace Security,Nanjing University of Information Science&Technology,Nanjing 210044,China)
出处
《电子测量技术》
北大核心
2024年第1期150-158,共9页
Electronic Measurement Technology
基金
国家自然科学基金(62276139,U2001211)项目资助。
