Investigation on Image Reconstruction Algorithms for Single Pixel Camera Using Compressive Sensing

Abstract

Compressive sensing has recently emerged as a front line research area in the field of signal processing and has attracted many researchers towards development of signal recovery algorithms. The compressive sensing theory provides a fundamentally new approach in image reconstruction using significantly lesser samples than Nyquist sampling criteria. . It shows that certain signal or image can be reconstructed exactly or sometime reconstruct perfectly from the incomplete data sets. This opened up a new pathway in digital camera technology, as exotic detectors with huge pixel banks can be avoided in the focal plane. The Digital Signal Processing group at Rice University created a singlepixel camera architecture based on compressive sensing technique. A single pixel camera compresses the signal while sensing, thereby greatly reducing the data throughput, scale, complexity and system cost. The Single pixel camera architecture calls for development of camera hardware and a suitable image reconstruction software module for efficient image reconstruction using the single pixel output. This thesis focuses on studying various algorithms available in the literature for image reconstruction efficiency in terms of various performance parameters, computation efficiency, robustness of the algorithm particularly for complex remote sensing images. In this thesis, initially, theoretical background and literature survey has been explained followed by current trend of Compressive sensing and Single Pixel Camera architecture. We tried to do systematic study of the available algorithms for image reconstruction to solve an under-determined system i.e. samples points are lesser than actual signal length. We tried to implement various algorithm namely OMP, CoSaMP, IHT and IRLS on few standard as well as typical remote sensing images of urban landscapes. Our study shows that IRLS algorithm provide better results for image quality which is depicted in terms of PSNR, MSE,SSIM for both image types. . In the process, a modified version of OMP algorithm has been attempted in which insisted of taking the random measurements, these has been controlled based on predefined DCT block size and zigzag scanning carried out.This method not only provides the good image quality but also the faster computation results then the conventional OMP process. Our study also concludes that the image reconstruction and its quality depends on degree of sparsity that an image have.