Research Of Computational Imaging Technology Based On Compressive Sensing

High spatial resolution, high spectral resolution and signal-to-noise ratio(SNR), high efficient image acquisition and transmission and storage technology are always the developmental goal of many kinds of optical imaging instruments. For the traditional information acquisition based on Shannon’s sampling theory, improving the resolution means increasing the number of photo-detector pixels and AD sampling frequency, which will dramatically increase the complexity of system and technical implementation difficulties. Moreover, in the field of military aerospace imaging, localization of large area array, infrared detector with high frame rate, and high-speed AD devices also have technical bottleneck, meanwhile these foreign advanced technology are prohibited to transport to China. Therefore, it’s important to develop all kinds of new type of imaging technology.The combination of computational imaging and compressed sensing theory(CS) provides a new solution. Unlike traditional optical imaging method for direct detection of image signals, this thesis proposes a new way of computational imaging using a special code template to replace the image plane of the traditional optical imaging and the coding template can be loaded with different coding function. And then CS is used in the computational imaging, the high-dimensional signal is projected onto another low-dimensional space by the observation matrix loaded to code template when the signal is sampled, which we can use a single-pixel detector or a low-speed AD to collect the projected signal. It can be seen that the new signal acquisition mode avoid the pursuit of large area array detectors and high speed AD. At the same time, signal in the sampling process achieves compression, which will greatly reduce the pressure of imaging image acquisition and transmission and storage for the application of aerospace remote sensing imaging.The main work and innovation of the thesis are as follows:1. The prototype of the computational imaging was designed. Using DMD as the encoding template hardware, the combinations and partitions of DMD’s array unit can be arbitrarily controlled based on the different size and location of the imaging target. At the same time, the coding template can be arbitrarily replaced for any noncompressed and compressed computational imaging experiments and the sampling signal are analyzed by statistics. Finally, multi-spectral imaging experiment of static target in the field environment was conducted based on the non-compressed computational imaging.2. Two types of representative compressed sensing signal reconstruction algorithms which contain the greedy algorithm and gradient projection algorithm had been well studied, then the OMP algorithm and GPSR algorithm were used to reconstruct the image. At last, the advantages and disadvantages of the two kinds of algorithms are analyzed in detail.3. In the process of the reconstructed image quality assessment, because of the limitations of the traditional mean square error and Peak-SNR method, we innovatively introduced the method of subspace analysis to estimate SNR. Then the estimated SNR is used to evaluate the quality of reconstructed image. This evaluation method agrees very well with experimental results.4. Finally, based on the research results, put forward some reasonable suggestions on the future application of the computational imaging technology. When you need to obtain high-resolution images, choose non-compressed computational imaging, and use Hadamard matrix as the encoding template. When the need for efficient during image acquisition, choose compressed computational imaging, use Gaussian random matrix as the encoding template, while select about 38% of the compressed signal sampling rate, and select the gradient projection algorithm to reconstruct the image; When you need fast imaging, and accepting a certain image resolution is reduced, you can use about 16% of the compressed signal sampling rate, and use orthogonal matching pursuit algorithm to reconstruct the image.