一种结合空间相似性和RPCA的高光谱图像去噪算法

A hyperspectral image denoising method using spatial similarity and RPCA

高光谱图像在采集过程中极易产生高斯、椒盐、条纹等噪声,从而对后续的地物空间识别工作产生影响.因此有效的噪声去除工作在高光谱图像处理中是不可缺少的一步.鲁棒主成分分析(Robust Principal Component Analysis,RPCA)是能将受稀疏噪声干扰的低秩矩阵进行有效恢复的模型.高光谱图像由于其光谱特征之间存在很高的相关性,即每个光谱特征可以用光谱端元的线性组合来表示,因此高光谱图像具有高度低秩性,从而RPCA算法能在高光谱图像去噪中取得显著的效果.结合高光谱图像空间邻域相似性和改进RPCA(Spatial Neighboring Similarity and Improve RPCA,S_IRPCA),提出一种新的高光谱图像去噪算法.算法在去除噪声的同时,更好的保留了细节信息.实验表明,算法与主流的低秩恢复算法相比,无论在主观视觉上还是在客观评价指标上,都做到了显著提升.

At present, hyperspectral image has attracted more and more attention because of its rich spectral and spatial information that can recognize the ground objects. But because the hyperspectral image is very easy to produce Gauss, salt, pepper and stripe noise in the process of collection, has effect on subsequent spatial identification. Therefore, effective noise removal is an indispensable step in the hyperspectral image processing. Robust principal component analysis（RPCA） is a model which can more effectively recover the low rank matrix from sparse noise interference. Because of the high correlation between spectral features of hyperspectral images,the hyperspectral image itself has a very low rank. The RPCA model used to denoise the original hybrid photographic image with noise has raised more and more attention. The relevant RPCA algorithm used in the hyperspeetral image denoising has achieved outstanding results. A new hyperspeetral image denoising algorithm is proposed, which combines the neighborhood similarity of hyperspectral spatial feature and improved RPCA. Firstly, the low rank property of spectral information accords with the RPCA model assumption. Considering the Gaussian noise, the constraint term added to the Gauss noise is based on RPCA,which can more effectively eliminate the mixed noise. Then, we add the spatial similarity of the pixels of the hyperspectral image, that is, the similarity between the pixels and the pixels in the N neighborhood are added to the improved robust principal component analysis model. After that,the model is conductive to recover and rebuild the impaired spatial information and can retain more spatial detailed information of hyperspectral image. Whether the subjective vision or the objective evaluation index, the experiments show that the proposed algorithm has achieved significant improvement compared with the mainstream low rank model algorithm.