![]() Among them, the FFDNet has achieved very competitive noise reduction performance. One type of the method is to learn segmented image priors during the truncated reasoning (Chen and Pock, 2016) another popular type of the methods is plain discriminative learning, such as the MLP (Burger et al., 2012), convolutional neural network-based DnCNN (Zhang et al., 2017) and FFDNet (Zhang et al., 2018). These methods aim to learn image prior and fast inference from a training set of degraded and clean image pairs. Inspired by such research achievement, in this paper, we use tensor Tucker decomposition (Wang et al., 2017) to simultaneously model the inherent tensor structure and low-rank prior in both spatial and spectral domains.Īs deep learning develops, lots of deep network based denoising methods have emerged in recent years. Recently, some studies in practice and theory have proven that direct tensor modeling technology is superior to matrix technology in processing high-order tensor data, e.g., Liu et al., 2012, Yuan and Zhang, 2016, Cao et al., 2016 and Anandkumar et al. However, such matricization usually fails to exploit the essential tensor structure, i.e., the spatial-and-spectral information in HSIs, resulting in poor recovery performance in some heavy noisy situations (Wang et al., 2017). Subsequently, 3-D HSIs are ordered into 2-D Casorati matrices whose columns comprise vectorized bands and use the low-rank prior of Casorati matrices to characterize the correlations between different bands (He et al., 2015b, Zhang et al., 2013). Although these methods can partly remove noise in HSIs, they ignore the correlations between different bands of HSIs. Therefore, as a pre-processing step, HSI restoration is an important research direction that needs to be further studied.Ī well-known method for HSI restoration tasks is to use grayscale image restoration methods to restore HSIs band by band (Dabov et al., 2007, Buades et al., 2005). These noises adversely affect the image quality of HSIs, the subsequent processing and applications, e.g., feature classification (Zhang et al., 2015), target detection (Stein et al., 2002), unmixing (Bioucas-Dias et al., 2012) and so on. However, due to the limitations of observation conditions and sensors, the HSI obtained by hyperspectral imagers is usually contaminated by a variety of noises, such as Gaussian noise, stripes, deadlines, and impulse noise. In recent years, they have attracted great research interest in the field of remote sensing. HSIs are widely used in many fields, such as environmental research, agriculture, military and geography (Goetz, 2009, Chan, 2019). Hyperspectral images (HSIs) can provide spectral information of hundreds of continuous bands in the same scene. ![]() Experiments with simulated data and real data show that, compared with competitive methods, the proposed method achieves better HSI restoration results in various quantitative evaluation indicators. The proposed model can be quickly solved using the alternating direction method of multipliers method. Then, we combine the advantages of the two methods to introduce the HSI restoration CNN with the low-rank tensor approximation based regularization in the flexible and extensible plug-and-play framework. And for the implicit prior, a CNN based method can represent the prior which cannot be designed by mathematical theory tools. For the physical prior of HSIs, a Tucker decomposition based low-rank tensor approximation can fully explore the global correlations in both the spatial and spectral domains. In this paper for HSIs restoration tasks, we firstly investigate the advantages of traditional physical restoration models and the denoising convolutional neural networks (CNN). ![]() However, during the imaging process, they are often contaminated by various noises. Hyperspectral images (HSIs) are widely used in various tasks such as mineral detection and food safety. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |