Distributed Resolution Enhancement Techniques for Remotely Sensed Image

Abstract

Super-Resolution (SR) refers to the reconstruction of a high-resolution image from one or more low-resolution images for the same scene. In this work, we presented enhancement and modification for example-based and compressed sensing super-resolution techniques for satellite images. The proposed techniques managed to solve the ill-posed inverse super-resolution problem and successfully managed to eliminate the disadvantages and solve the existing problems related to the classical bicubic interpolation and the state-of-the-art super-resolution methods such as; POCS, IBP, Freeman method, and Chang method. The proposed techniques were tested and evaluated against several datasets. The results showed improvements in the quality, and robustness of the high-resolution results compared to the classical bicubic interpolation and other related state-of-the-art techniques. Furthermore, extensions to compressed sensing based super-resolution methods are provided. Another extend to Self-similarity method was also presented and evaluated against the the classical bicubic interpolation. In this Chapter, we conclude our work, a summary of our proposed work and the experimental results are presented in the first section. Directions of the future research are presented in the second section. 7.1 Conclusion 7.1.1 Super-resolution using support vector machine Example-based methods are used to break the limitation of classical SR methods. The learning-based techniques are classified into supervised learning and reinforcement learning according to the model used. The statistical learning is also known as a machine learning-based method; bring prior knowledge in the reconstruction process. The goal of these methods is to learn the mapping between low and high images then the learned mapping is adopted to obtain a high-resolution image when given a low-resolution image. The learning-based scheme outperforms the classical methods even if the number of low-resolution images are reduced to one. Chapter 4 Section 4.1.1 presented an enhancement of super-resolution based support vector regression for satellite images. The modified technique combined with the PCA to reduce the dimensionality using the magnitude value of vertical and horizontal gradient for patches of the image. In spite of improvement the reconstructed high-resolution image, the need for real-time increased. Chapter 4 Section 4.1.3 introduced a parallel implementation to the proposed super-resolution technique based on support vector regression. CUDA, NVIDIA's parallel computing architecture is used to implement kernels for both training and testing phases.The artifacts and the blurring in the reconstructed high-resolution images using the proposed technique was reduced with acceptable accuracy compared with the state-of-the-art methods. The proposed GPU technique using support vector regression with optimal parameters (c=260,ε=1,σ=2)outperformed the classical bicubic interpolation in terms of PNSR values. The image quality increased from average PNSR 26.44 up to 28.29 and 26.44 up to 28.29 for systaltic and multispectral dataset respectively. In terms of patch size, average PNSR for 3×3 and 5×5 patch sizes archived 26.788 and 27.134 respectively then increased remarkably to reach the highest point with average 27.314 using patch size 7×7. At patch size 9×9 and15×15, the PNSR values dipped to average 26.588 and 23.453 respectively. In terms of sensitivity to upscaling factor, the average quality of the reconstructed image decreased from 27.52 to 23.808 for upscaling from 2× up to 5×. In terms of speedup, the proposed GPU implementation archived 50× over CPU serial implementation in the training phase and 27× for the reconstructing phase. 7.1.2 Super-resolution using manifold learning Neighbor Embedding (NE) is one of the powerful manifold techniques used in super- resolution, but due to the complexity of manifold learning algorithms, their vast computation times are very challenging. In Chapter 4 Section 4.2.2,an example-based super-resolution method for satellite images based on LLE algorithm is proposed. The Locally Linear Embedding (LLE) algorithm requires massive computing power, especially when applied to satellite images. An acceleration of the locally linear embedding algorithm was introduced in Chapter 4 Section 4.2.3 to be employed in an example-based super-resolution method. The algorithm is implemented using CUDA API and the performance is evaluated and compared with the implementation of the same algorithm on a single CPU and the state-of-the-art algorithm such as; bicubic interpolation algorithm and nearest neighbor embedding algorithm. The proposed super- resolution techniques based manifold reduced the artifacts and blurring in the resulted high-resolution image. The proposed CUDA implementation for the manifold learning model outperformed the classical bicubic interpolation in terms of PNSR values. The image quality increased from average PNSR 27.087 up to 27.515 and 26.44 up to 28.29 for systaltic and multispectral dataset respectively. In terms of sensitivity noise, all PNSR values for all SR algorithms declined over the increase of the noise. In the bicubic and the Chang methods, the quality of the HR image changes marginally at the noise variance (σ = 5, σ = 10, σ = 20), while the quality decreased slightly from 2.6% at σ = 20 to 8.6%at σ = 50 for the bi-cubic method and from 6.8% at σ = 20 to 10% σ = 50 for the Chang method. In the Freeman method, the quality of the recovered image declined dramatically from 8.7 to 10.1% for noise variance σ = 20 and σ = 50 respectively. In the proposed method, the PNSR values are regularly but only slightly change from one noise variance to another. The PNSR values dropped by 7.9% at σ = 50. Overall, the proposed method outperforms other algorithms in terms of PNSR. In terms of higher upscaling factor, the average quality of the reconstructed image decreased from 28.9, 28.2, 27.5 and 24.394 for upscaling from 2× up to 5×. In terms of the speedup, in terms of acceleration, the proposed GPU implementation enhanced speed-up compared with the sequential CPU, from 10.20 for small images (64×64) to 162.90 for large images (1024×1024).