부산대학교 도서관

Hyperspectral imagery is an emerging technology that can simultaneously take images of the same scene in many contiguous and narrow spectral bands, and it can acquire a faithful representation of the scene radiance by a distribution of intensity at each wavelength. The captured hyperspectral image (HSI) is a spatial-spectral merging three-dimensional data cube that contains both spatial and spectral information. However, compared with RGB images and 3-D voxelized objects, it is a challenging problem to model the abundant spatial-spectral information of HSI cube.Deep learning, especially the discriminative model based on convolutional neural network for image analysis, has shown great potential in HSIs. Using 3-D filters is flexibly capable of receiving HSIs with any number of bands and can extract the feature maps along both spatial and spectral dimensions simultaneously. Therefore, 3-D convolution is well-suited for various HSI task. For HSI restoration, it is crucial to extract more context information around each pixel and to predict each pixel according to the surrounding context. Therefore, the effective receptive field plays an important role in performing HSIs restoration tasks. Generally, the HSI restoration model can achieve better performance by reserving the correlation of adjacent spectral bands and extracting more pixel features in the spatial domain. In this dissertation, two 3-D atrous convolution neural networks are proposed for HSI denoising and super-resolution. The models extract features along both spatial and spectral dimensions and enlarge the receptive field. Simultaneously, the multi-branch and multi-scale structure is utilized to reduce training difficulty, lessen overfitting risk, and preserve details in texture. The experimental results of the quantitative and qualitative evaluation show that our proposed methods yield satisfactory performance.Next, we explore a learning data augmentation method for hyperspectral image classification with adversarial networks. A new multi-stage and multi-pole generative adversarial network is proposed, which is suitable for conditional hyperspectral image generation and classification (HSIGAN). For HSIs synthesis, it is crucial to learn a great deal of spatial-spectral distribution features from source data. The multi-stage and progressive training makes the generator to effectively imitate the real data by fully exploiting the high dimension learning capability of GAN models. The information extraction of coarse-to-fine helps the discriminator to understand the semantic feature better while the integrated prediction presents a positive impact on HSIs classification. A spectral classifier joins the opposing camps of the original bi-polar adversarial network as a new pole and then offers a helping hand to stabilize and optimize the model. Moreover, we apply the 3D DropBlock layer in the generator to remove semantic information in a contiguous spatial-spectral region and avoid model collapse. Experimental results of the quantitative and qualitative evaluation show that HSIGAN could generate high-fidelity, diverse hyperspectral cubes while achieving top-ranking accuracy for supervised classification. This result is encouraging for using GANs as a data augmentation strategy in the hyperspectral image vision task.Finally, the dissertation proposes to produce a classification-driven HSI denoiser that is capable of simultaneously reducing noise and preserving semantic-aware detail for the high-level vision tasks. The conditional neural adversarial network is introduced to produce visually pleasing images by enforcing an additional constraint that the denoised image must be indistinguishable from its corresponding ground-truth clean image. The proposed model copes with both low-level denoising and high-level classification tasks jointly and explores the mutual influence between them to achieve mutual benefits. A denoiser is trained to transform noise HSIs into routine clean images by using hybrid losses which include pixel loss, adversarial loss, and feature matching loss. A multi-task and multi-scale discriminator is simultaneously trained to give both a probability distribution over sources and a probability distribution over the class labels. Experiments are performed on hyperspectral remote sensing images containing both the simulated Gaussian noise and real noise. The results show that the proposed model outperforms many state-of-the-art denoising and classification methods for HSIs.