Computational Spectral Imaging Based On Compressive Encoding

The spectral images of remote sensing take great advantages over the identification of objects in the scene, so that they have achieved extensive attention in the field of remote sensing geological survey, forest monitoring, agricultural applications, and marine research, and so on. A spectral image with higher resolution means it has more detail about the objects, thus spectral resolution is one of the important parameters of the spectral imager. However it is limited by device size, quality, power consumption, transmission bandwidth and other conditions to improve the resolution by improving the sensor density. It becomes a challenge that how to break through such limitations and to fulfill these needs in the field of spectral data application. In recent years, computational spectral imaging methods have been developed to provide a new way for the design of imaging spectrometry. In such research, the optical elements are taken as computational components for realizing the special relationship between spectral signal of the scene and the sensor array on the imaging plane, and then spectral data cube can be acquired by a reconstruction algorithm.Based on the theory of compressive sensing (CS), this dissertation studies the problem of signal sampling in spectral imaging system, and proposes a method of computational spectral imaging method based upon compressive encoding. The computational spectral imaging method is divided into two steps:encoding perception of the scene and the reconstruction of spectral data, that is corresponding to the two steps of the CS, the compressive measurement and the reconstruction by optimization algorithm. The computational spectral imaging method has been developed with the aid of theoretical foundation of CS, such as measurement matrix, sparse representation, and reconstruction algorithm.The main content of this dissertation is to realize the "compressive encoding" in the process of imaging, and design the corresponding computational spectral imaging method, the work and innovation are described as follows:1. A multi-slit push-broom spectral imaging method is provided. The acquisition of high image quality and high resolution spectral data is limited by light flux. A push-broom spectral imaging should reduce the spatial resolution if it amplified its light flux to increase its signal noise ratio. According to this problem, the theory of CS is introduced for modeling push-broom spectral imaging system from the signal processing analysis, so that the number of slit of the imaging system can be increased to amplify its light flux. The light flux can increase without reducing the spatial resolution. In the simulation, its exposure frequency dropped to 1/4 of the original, and its light flux increased to 128 of the original, spectral image with resolution of 512x512 can be well obtained. It is suitable for remote sensing by using less times for imaging and less memory for storage and transmission compared with traditional one.2. A high-resolution spectral imaging method based on coded dispersion is provided. Since the energy of the incident light is constant, the spatial and spectral resolution can hardly be improved without scarifying the other with spectral imaging method of pushbroom scanner. Thus, a new spectral imaging method is proposed to obtain high-resolution spectral image with low-resolution detector array by improving the information collection efficiency of imaging system. Two modes for realizing coded dispersion in the step of compressive measurement are provided:revolving dispersive element and coded diffraction grating, and then reconstructed high-resolution spectral images are obtained by optimization algorithm. The simulation result shows that the proposed method also offers a new way to acquire high-resolution spectral image with low-resolution sensors.3. A compressive spectral imaging method using variable number of measurements is provided. CS breaks through the limitation of Nyquist’s theorem, and enables us achieving the information of object by fewer samples. Our method cuts down on the number of sensors on the imaging plane, so as to fit some practical constraints. The proposed method is based on the concept of coded dispersion. Its measurement matrix is modified so that the number of measurements can be variable under different circumstances to save the transmission bandwidth. Moreover, fewer samples are needed, while more prior knowledge about the object in the scene has been achieved. Thus, it provides us an approach to balance between the two demands that high image quality and shot transmission time. Some simulations have been designed and the results demonstrate the validity of the proposed method.