Research On Single-pixel Compressive Imaging Based On Convolutional Neural Network

Recently,with development of information technology,imaging device is developing for miniaturization and high quality.The computational imaging theory that replaces the main components of the imaging system by calculation has become a research hotspot in the fields of modern optics,image processing,machine learning,etc.Compressive imaging is an important research in the field of computational imaging.Compressive imaging originated from the theory of compressed sensing,is a method to integrate compressed calculation into optical system design.Compressive imaging through the observation matrix to project high-dimensional signals into lowdimensional space during image acquisition,and the reconstructs the original signals from low-dimensional observation data.The method of compressive imaging can effectively reduce the amount of data and is of great significance for the acquisition,storage and transmission of large data images.Compressive imaging mainly contains the design of observation matrix and image reconstruction.In this thesis we mainly focus on the hardware implemented of image reconstruction.In recent years,deep neural networks have been extensively studied in the field of image recognition and achieved great success,but research on image reconstruction has just begun.There is a many-to-one relationship between input data and output data in image recognition and classification tasks.However,there is a manyto-many relationship between input data and output data in image construction tasks,so the image reconstruction task will contain more parameters.In general,the number of parameters required for an image reconstruction system using the same structure is1.5 to 1.8 times that of the recognition system.For example,the residual image reconstruction network needs 43 million parameters,and the image identification network only contain 25 million parameters.So,during hardware implementation of image reconstruction algorithm based on neural network,the insufficient resource of memory will become the main bottleneck.In this thesis,we propose the structure of the ordinary differential equations(ODE)network,compresses the neural network hierarchy with ODE,and reduces the amount of network parameters.Firstly,we study the image reconstruction neural network using domain-transform manifold learning.Then the basic principle of model compression is introduced.The ODE solver is applied to the neural network to construct the ODE neural network,which replaces the calculation between discrete layers in the traditional neural network.Through the MNIST handwriting recognition task comparison the residual network and the ODE neural network,it can be proved that the parameter quantity of the ODE neural network does not increase with the increase of the number of evaluations,and compress about 20% of network parameters relative to the residual network,and the accuracy of the ODE network can be improved by about 0.5%.Subsequently,the ODE neural network is applied to the image reconstruction neural network.Comparing the 64×64 input images,the experimental results show that the ODE image reconstruction neural network can achieve the similar effect as the domaintransformation manifold learning image reconstruction network,and can reduce17,039,360 parameters.Afterward,we propose the hardware optimization strategy of each computing module in the ODE image reconstruction neural network and quantize the data.After the optimization strategy,the ODE image reconstruction neural network can save 170 MB storage space than the domain-transformation manifold learning image reconstruction network.Finally,ODE image reconstruction neural network is implemented on the FPGA Zynq ZCU102 development platform of Xilinx.