Imaging and Visual Classification by Knowledge-Enhanced Compressive Imaging

Compressive imaging is a technology that uses multiplexed measurements and the sparsity of many natural images to efficiently capture and reconstruct images. The compressive single pixel camera is one embodiment of such an imaging system and has proven capable of imaging static images, dynamic scenes, and entire hyperspectral datacubes using fewer measurements than the current schemes. However, for many imaging tasks prior information or models exists and when incorporated in the compressive measurement can greatly improve reconstructed result. In this thesis, we illustrate and quantify through simulation and experiment the effectiveness of knowledge-enhanced patterns over unbiased compressive measurements in a variety of applications including motion tracking, anomaly detection, and object recognition. In the case of motion tracking, one might interest in moving foreground. Given prior information about the moving foreground in the scene, we propose the design of patterns for foreground imaging. Then one can recover the dynamic scene through combining moving foreground from designed patterns and static background. We also implemented anomaly detection from compressive measurements. A set of detection criteria is implemented and proven to be effective. On top of that, we also introduced patterns selected from partial-complete set according to the geometric information of the anomaly point, which later shows improved effectiveness comparing with random patterns. For image classification, we implemented two methods to generate secant projections, which are optimized to preserve the difference between image classes. Lastly we illustrate the new design of single pixel based hyperspectral design. To reach that, the control of DMD chip and optics of SPC have been improved. Also we show results about implementation of compressive endmembers unmixing scheme for compressive sum frequency generation hyperspectral imaging system