Learning Deep Dictionary for Hyperspectral Image Denoising

Abstract

SUMMARYUsing traditional single-layer dictionary learning methods, it is difficult to reveal the complex structures hidden in the hyperspectral images. Motivated by deep learning technique, a deep dictionary learning approach is proposed for hyperspectral image denoising, which consists of hierarchical dictionary learning, feature denoising and fine-tuning. Hierarchical dictionary learning is helpful for uncovering the hidden factors in the spectral dimension, and fine-tuning is beneficial for preserving the spectral structure. Experiments demonstrate the effectiveness of the proposed approach. key words: dictionary learning, deep learning, hyperspectral image denoising, fine-tuning

Hierarchical Sparse Learning with Spectral-Spatial Information for Hyperspectral Imagery Denoising

“…The compared methods include K-SVD [9], BM3D [10], ANLM3D [13], BM4D [19], LRMR [31] and DDL3+FT [34]. The necessary parameters in the K-SVD, BM3D, ANLM3D and DDL3+FT methods, were finely tuned or selected automatically to generate the optimal experimental results as the reference suggested.…”

“…Zhu et al presented an HSI mixed-noise removal method [32], which achieves the promising image quality by combining low-rank constraint and total-variation-regularized. Deep learning has been widely used in the HSI analysis [33,34]. Motivated by this, a deep dictionary learning method (DDL3+FT), consisting of the hierarchical dictionary, feature denoising and fine-turning, is developed to effectively suppress the noise in HSI [34].…”

“…Deep learning has been widely used in the HSI analysis [33,34]. Motivated by this, a deep dictionary learning method (DDL3+FT), consisting of the hierarchical dictionary, feature denoising and fine-turning, is developed to effectively suppress the noise in HSI [34]. …”

Image Pattern Similarity Index and Its Application to Task-Specific Transfer Learning

“…In this part, the sparse representation [9], which involves generating common dictionary with training images by optimizing a pre-defined objective function followed by reconstructing testing images with the dictionary, is introduced for figuring out the shared patterns by designing a structural sparse dictionary, where the atoms representing the similarity of input images are obtained. The atoms to dictionary is what the words to sentence, and the atoms are usually generated by clustering algorithm, and dictionary is just a codebook by grouping the learnt atoms., and the number of the shared atoms is the objective quantized index for evaluating the shared patterns between different images.…”

Hierarchical Sparse Bayesian Learning with Beta Process Priors for Hyperspectral Imagery Restoration

“…K-means and Ksingular value decomposition (KSVD) [13] adopts the local sparseness characteristic to remove additive Gaussian noise. Deep Dictionary Learning is explored in [14]. Due to the high correlation between hyperspectral signatures of pixels, the low rank constraint is incorporated as a regularization term to denoise the HSI [15].…”