Research On The Key Technologies Of Compressive Sensing

The compressive sensing(CS) theroy pointed out that for sparse or compressiblesignals, sampling rate is determined by the structure and content of information insignal rather than by the bandwidth of signal. Currently, much achievement has beenyielded in the research on CS theory and applications, while there are still many openproblems to be studied further. For the existing proplems,this disseration studied somekey issues as directional transform, feature dictionary, video analysis, compressiveimaging which are theoretically important and practically valuable. The maincontributions and innovations are as the following:Firstly, the practice of directional-transforms method in CS construction of imagewas studied which using a general frame of projected-based reconstruction withblocked random sampling of images to improve sparsity and smoothness as well. Thisframework facilitated introducing of directional transforms based on contourlets anddual-tree complex wavelets into the CS reconstruction and finally led to the fastfunction speed of the projection-based CS reconstruction. Simultaneouly, thecombination of a smoothing step and boosted directionality naturally improved imagequality, particularly for the cases at low sampling rates.Secondly, taking advantage of the sparsity of Hough transform, a sparsifyrepresentation dictionary of shapes was built which using CS method to search forparameterized shapes in images. An experiement on detecting lines and circles in annoisy image from a few CS measurements was implemented indicating that there’spossibility to get cleaner results than Hough transform. Furthermore, the analysis ofdetect rate v.s. SNR and detect rate v.s. number of measurements was conducted.Thirdly, a privacy-protection video-monitoring encoding scheme was designedwhich was able to track an object without the need to reconstruct each original frame.Subsequently, this scheme ensures to encoding a video sequence by using a fewpesudo-random projections of each frame, while the decoder reconstructs the positionof foreground targets by exploiting the sparsity of background-subtracted images. Itworks with a particle filter in order to estimate the position of targets, then theestimated position in turn serves as a prior to promote the recovering of foreground.Privacy protection roots in that it is impossible to reconstruct the original content only with a coded random projection, meanwhile, security roots in that if there is no seedwhich creates random projection, it would never decode.Last but not the least, CCM used DMD to efficiently scan2D or3D specimenand all the resulting data coming from a random collect of pin-hole illuminatedspecimen pixels, measured by a single photon detector after linear combination(projection). Compare this with the traditional methods of CM or PAM, CCM speaksfor itself for its capacity of simplifying the complexity of confocol imaging hardwareand optics by off load processing from data aqusition stage to software image recoverand getting more cost-effective. Besides, CCM also provides special opticalsectioning property of confocol imaging at reduced sampling rate.3D jointlyreconstruct approach fully utilized the correlation among3D slices to promote systemperformace. In a word,3D method significantly increased the PSNR of imageconstruction while kept the same computational complexity.