A Study On Compressive Hyperspectral Imaging Reconstruction Algorithms

Hyperspectral images extend the human visual power from the visible spectrum to a wider wave band,thereby enhancing human's observation ability.For example,for two RGB images of the same color,if their spatial distribution are similar,we may not be able to distinguish between substances in them by the naked eye alone.However,finer spectrum can help us to distinguish different materials better.This is because when the electromagnetic wave shoot onto the surface of the material,the transition of electrons,rotation of atoms and molecules within the material at a particular wavelength would form a unique absorption and reflection characteristics.Therefore,from the opposite perspective,we can spy on the material properties from the material absorption and reflection characteristics.Since the spectral image is three-dimensional,and our detectors are up to two dimensional array detector,the traditional spectral imager have to get multiple exposure of a scene,in order to detect the whole three-dimensional images.This greatly limits the temporal resolution of the spectral imaging instrument,especially when there are moving objects in the scene,we can not do real-time target detection.The theory of compressive sensing has brought a breakthrough to this limitation.Based on this theory,Duke University's DISP group developed a coded aperture snapshot spectral imager(CASSI).The three-dimensional spectral data cube is encoded,and then the two-dimensional array is used to detect the encoded signal.Finally,the original three-dimensional spectral image can be reconstructed using the proper reconstruction algorithm.This greatly improves the temporal resolution of spectral imageing system.We should solve a system of linear equations to recover the original datacube,and according to the theory of linear algebra,we know that it exists a solution space.In order to get a unique solution,the sparseness of the natural signal becomes the key.Compressive sensing reconstruction algorithms have been developed for a long time from the initial greedy algorithms to the recent data-driven methods based on convolutional neural networks,which brings about the improvement of reconstruction performance and the speed of reconstruction.In the third chapter,we systematically combs the compressive reconstruction algorithms for CASSI system.It is a holistic system from the initial signal acquisition,to image reconstruction,and then to low-level visual processing,including noise reduction,de-blur,etc.,and finally to the final interpretation of the image,including classification.The elements in the system are interrelated,especially,some elements are overlapped.According to this idea,in the second section of the fourth chapter,we merge the super-resolution process into the imaging hardware.So the original two steps are simplified as a step by making the compressive reconstruction algorithm and super-resolution combined together.Also,we achieved high-quality reconstruction images.At the same time,in the first section of the fourth chapter,we combined hyperspectral image compressive sensing reconstruction step and unmixing task.So that the original two tasks can be completed at the same time,reducing the redundancy in the system and the complexity of the entire system.Also,we get satisfactory results.