基于非局部低秩和全变分的高光谱图像去条带方法

Hyperspectral Image Destriping Method Based on Nonlocal Low-Rank and Total Variation

受探测器的像素响应不均匀、传感器的机械运动以及图像采集过程中温度的变化等因素的影响,获取的高光谱图像中常包含条带噪声,而当前去条带方法往往着眼于条带的整体性质而忽略了条带的非局部相似性,难以取得较满意的去条带结果。针对上述问题,文章通过分析条带噪声和干净图像的先验信息,提出了一种基于非局部低秩张量分解和全变分的去条带算法。该算法考虑了条带的非局部相似性,对与参考块相似的条带进行聚类,进而用张量低秩分解进行逼近,另外,还考虑了条带的方向和结构稀疏特征,通过联合高光谱图像的局部和非局部相似性,实现光谱畸变的有效降低。为评估该方法的去条带效果,分别进行了模拟数据试验和真实数据试验,其中模拟数据试验结果显示,在随机长度条带和整体条带情况下,该算法的平均峰值信噪比(MPSNR)和平均结构相似性指数(MSSIM)值分别比对比方法中最好的结果高出约2~3 dB和0.02~0.04,而平均光谱角匹配(MSAM)值降低约0.02~0.06;真实数据试验结果显示,该算法能够精确地估计和分离条带,恢复出受条带影响的图像信息,克服条带残留问题,并且其无参考评价指标逆方差系数(ICV)和平均相对偏差(MRD)均优于对比方法。文章提出的算法为去除高光谱图像中的条带噪声提供了一种有效的解决方案,有望为高光谱图像的后续应用提供有力的支撑。

Due to factors such as uneven pixel response of the detector,mechanical motion of the sensor,and temperature changes during image acquisition,hyperspectral images often contain stripe noise.Current destriping methods often focus on the overall properties of the stripes and ignore their non-local similarity,making it difficult to achieve satisfactory destriping results.To address the above issues,this article proposes a destriping algorithm based on non-local low-rank tensor decomposition and total variation by analyzing the prior information of stripe noise and clean images.This algorithm considers the non-local similarity of stripes,clusters stripes similar to the reference block,and then approximates them using tensor low-rank decomposition.In addition,it also considers the directional and structural sparse characteristics of bandstripes,and achieves effective reduction of spectral distortion by jointly considering the local and non-local similarity of hyperspectral images.To evaluate the destriping effect of this method,we conducted both simulated data experiments and real data(Data from the Gaofen-5 satellite and data captured by the EO-1 Hyperion hyperspectral sensor of the Earth observation satellite in a certain region of Australia)experiments.The results of the simulated data experiments showed that under random length stripes and overall stripes,the mean peak signal-to-noise ratio(MPSNR)and mean structural similarity index measure(MSSIM)values of this algorithm were about 2~3 dB and 0.02~0.04 higher than the best results in the comparison method,respectively,while the mean spectral angle mapper(MSAM)value decreased by about 0.02~0.06.The results of real data experiments show that the algorithm can accurately estimate and separate stripes,recover image information affected by stripes,overcome the problem of residual stripes,and outperform the comparison method in terms of the inverse coefficient of variation(ICV)and mean relative deviation(MRD),which are non-reference evaluation metrics.The algorithm proposed in this article provides an effective solution for removing stripe noise in hyperspectral images,and is expected to provide strong support for the subsequent applications of hyperspectral images.