Research On Fast Imaging Of Single Pixel Camera

As a new direction in the field of image processing,compressive sensing theory has been widely and deeply studied during the past years.The single pixel imaging has become a hot topic as it is an extension of compressive sensing.as its computational photography model surpasses the traditional one in imaging,thus attracting the attention of reserachers.The key issue of single pixel imaging system is how to effectively improve the quality and speed of imaging.Focus on these two aspects,this paper makes an intensive study on measurement matrix designing and signal reconstruction algorithm and proposes some strategies to improve its performance.The main work and innovative points are included as following:(1)Measurement matrix generation method using hierarchical model and guide prior.Based on the relationship between measurement matrix and image signal,this paper proposed a hierarchical model and guide prior constraint to improve the image quality and imaging speed of the single pixel imaging system.Firstly,by analyzing the relationship between different resolution scale,a hierarchical model is established to depict such relationship.Secondly,by using the prior knowledge of structure and location information,the mutual use of inter layer information in hierarchical model is realized.Finally,the designing method of measurement matrix in single pixel imaging system is improved by ultilizing the proposed model.Experimentas have shown that the improved method based on the hierarchical model and guide prior generates better reconstruction results with faster imaging speed.(2)Measurement matrix generation method using feedback model.From the observation that the non-zero items of vast majority of sparse signals are not randomly distributed but reflect certain structural characteristics,this paper proposes a feedback model to generate measurement matrix of single pixel imaging system by using the constraint extracted from the generated measurement matrix sequence.First of all,from the perspective of Bayesian theory,the rule is figured out that measurement matrix with large amplitude can be used to accelerate the convergence speed of the reconstruction algorithm.Then the feedback model of the measurement matrix is established to constrain the generation of matrics.Experimentals have shown that the proposed feedback model can effectively reduce the searching range of the feasible solution in the process of signal reconstruction and accelerate the convergence speed.Meanwhile,the proposed method shows good robustness against to noise.(3)Video generation method of single pixel imaging system using group constraint.Aiming at generating video via single pixel imaging system,this paper proposes a spatial group constraint based method and conductes a preliminary exploration of relevant video generation.First of all,this paper establishes a framework to generate video frames via single pixel imaging system by adding frame difference.The advantage of this framework is the reduction of data.Then by analyzing the frame difference signal of continuous video frames,it is found that frame difference has sparse features and spatial group characteristics.Finally,a fast frame difference signal reconstruction algorithm based on sparse and group is proposed.Experimentals have shown that the proposed method accelerates the reconstruction process while still guarantees the quality of signal reconstruction.(4)Signal reconstruction of single pixel imaging system using DCNN based method.In order to further accelerate the reconstruction algorithm and discard iterative scheme,this paper proposes a very fast signal reconstruction method based on deep convolutional neural network.The algorithm remarkably reduces reconstruction time by 3 or 4 orders of magnitude while ensuring the quality of reconstructed signal.The method constructs a two stage deep convolution neural network to reconstruct signal,and introduces a new loss function to improve the accuracy of image reconstruction.Experiments show that the network model trained by the ideal data set can not only reconstruct the high quality signal quickly,but also show good robustness against noise.At the same time,the same network model can be used to reconstruct the color image signals.