Research On Reconstruction Algorithm Of Deep Convolutional Network For Lensless Coded-mask Imaging

As a new type of imaging technology,the lensless computational imaging technology is based on the idea of replacing the lens group with optical modulation devices in front of the image sensor,achieve imaging through a thin lensless camera cooperating with the reconstruction algorithm.Lensless computational imaging technology has broad application prospects in Io T imaging scenarios requiring large scale deployment.Given the simple manufacture,low cost,and ease of integration of the coded mask methodology,the purpose of this thesis is to use the coded mask to replace the lens set and construct a lighter,thinner,and more affordable imaging system.Up to now,the imaging quality of the lensless imaging system based on codedmask still cannot match the standard of the traditional lens imaging,partly because the supporting image reconstruction algorithm is not yet mature.This thesis focuses on the reconstruction algorithm of deep convolutional network for lensless coded-mask imaging,and the main research achievements are as follows:(1)Targeting the research on the reconstruction algorithm of lensless imaging,a lensless imaging system based on the coded-mask is built,and a lensless coded-mask combining the TSVD-Tikhonov regularization and block matching 3D algorithm is proposed.The coded-mask was directly placed in front of the CMOS image sensor for several image calibration experiments.The algorithm was used to collect and reconstruct images of different scenes,and a number of objective image evaluation indexes were used to comprehensively evaluate the reconstruction results.The experimental results show that the algorithm proposed in this thesis performs well in the visual effect and objective evaluation index of the image acquired by the lens-less coding mask imaging system for image reconstruction.(2)To meet the requirements of lighter and thinner imaging equipment in the future,the prototype of the lensless imaging system was optimized,and a lensless coded-mask imaging system based on a miniaturized multi-spectral CMOS image sensor was built.The miniaturized lensless imaging system is used for data collection,images reconstruction,and further analysis.(3)Faced with the defect that the ill-posed inverse problem in the process of image reconstruction by lensless imaging technology leads to the imperfection of scene image reconstruction,in the research work of this thesis,a lensless coded-mask computational imaging method combining deep convolutional network processing algorithm is proposed,the powerful feature extraction ability and the learning ability of complex mapping of deep neural network are used to improve the quality of reconstructed images.The imaging method proposed in this thesis is used to collect images of different scenes,solve the system transmission matrix for preliminary reconstruction of images,and performed deep convolutional neural network processing.The output image results of different algorithms were obtained and analyzed,and the structure of deep convolutional neural network was improved to construct a Dense-U-Net network with both end-to-end I/O characteristics and dense connection characteristics of the same size.The processing of the reconstructed image using the network proposed in this thesis greatly improves the quality of the reconstructed image of the lensless imaging system.(4)According to the application requirements of lensless computational imaging technology in large-scale Internet of Things deployment scenarios,experiments were designed to explore the performance of the image reconstructed by the algorithm proposed in this thesis in image classification and image recognition tasks.The selftraining image classification and pre-training image recognition experiments were carried out on the image data set collected by the lensless imaging system based on coded-mask and output by image reconstruction and deep neural network processing.The processing accuracy of the image data set collected,processed and finally output by the lens-less imaging system based on coding mask in the most extensive image classification and image recognition tasks in the field of machine vision is studied.Finally,good results are obtained in the self-training image classification task and the pre-training image recognition task.