Adaptive Optimized Sparse Representation Based Compressed Sensing Reconstruction For Remote Sensing Images

Currently, almost all the optical remote sensing imaging systems are based on the Nyquist-Shannon sampling theory which notes that the sampling rate must be more than twice of the signal bandwidth in order to achieve accurate reconstruction. However, as higher resolution (HR) images are highly desired, which means imaging systems should have longer focal length and larger aperture, higher sampling rate focal plane arrays (FPAs) and smaller pixels, and these will greatly increase the designing and manufacturing difficulty of the FPAs and optical imaging system. The compressed sensing theory has noted that if the signal is sparse or compressible in some transform space, the signal can be reconstructed in high precision with little sampling rate great less than the Nyquist-Shannon Sampling Theorem required. Therefore, the use of compressive sampling theory in remote sensing imaging system can compress images while sampling, this is speedup and can save time, FPAs and memory space of the satellite. Meanwhile, adaptively selecting the optimal sparse representation method for the remote sensing images of different types can effectively improve remote sensing images’reconstruction quality and facilitate to the post information extraction. Besides, regardless of the observed image type, with fixed observation method can still get high quality reconstructed image.This paper first introduced the background and significance of our research and summarized the common image sparse representation and random sampling methods and whose corresponding observation matrix construction methods under the theoretical framework of compressed sensing. Further, we also described the common optimization reconstruction algorithms in compressed sensing and on the basis, we put forward some innovation methods and improvements in the field of image restoration and image fusion.Then, this paper focused on the training of over-complete dictionary with the use of K-SVD method and intensively analyzed the sparse representation performance of the training dictionaries for the different types of remote sensing images. We gave the details of the joint optimization method which is by taking iterative optimization on the observation matrix with a given training dictionary. We have finished the simulation experiments by adopting this method and obtain the optimization observation matrix and tested the reconstruction performance by the use of optimization observation matrix and training dictionary pairs. On this basis, we analyzed and summarized how the ratio between the reconstruction sparsity and training sparsity factor affects the image reconstruction and proceeded to propose that, by randomly selecting small amount of samplings which are corresponding to the original image blocks’measurements to reconstruct, the reconstruction property can help us estimate the better reconstruction sparsity. We also discussed the role and significance of image quality from subjective and objective evaluation aspects.Finally, We proposed the adaptive optimized sparse representation based compressed sensing reconstruction for remote sensing images. We successfully train and obtain three optimized dictionaries which are corresponding to the three kinds of remote sensing images including the cities, mountains and harbors respectively and we compare the three dictionaries with the DCT in sparse representation accuracy. Our compressed sensing reconstruction method have several steps. First, the preliminary identification of remote sensing image type can be judged from the crude compressed sensing recovery of the measurements. Next, we choose the corresponding optimized dictionary as the sparse representation method according to the image type. Then the accurate reconstruction algorithm is adopted to achieve the precise results. We main employ the peak signal to noise ratio(PSNR) and the structural similarity(SSIM) as the objective evaluation criteria and combine with the subjective evaluation to assess the reconstruction images. The experiment results show that our method can achieve the better image properties in compressed sensing image reconstruction.