Research Of Noise Elimination And Image Reconstruction For Quanta Image Sensor

The quanta image sensor(QIS)has three characteristics including spatial oversampling,temporal oversampling and single photon counting to deal with the drop of dynamic range and signal-to-noise ratio in the CMOS image sensor due to pixel shrink and the decrease of supply voltage.The QIS has a wide-ranging application space: low light imaging,high dynamic range imaging,high speed imaging and high resolution imaging,etc.However,there are still many problems to be solved in the research of QIS,such as the high-speed reconstruction of binary data and the efficient algorithm for noise elimination.Through the detailed theoretical and modeling analysis in this paper,the image reconstruction and noise elimination methods of QIS based on different readout architectures are designed.In this paper,a mathematical model is first established for the oversampling mechanism of QIS,and the working principle is analyzed as well as the important physical parameters.Then two readout architectures are researched: column readout and 3-D stacked parallel readout.Combined with the non-ideal factors,noise analysis is introduced.Aiming at the noise characteristics of column readout architecture,a noise elimination method based on the estimation of column noise vector by guide filter is proposed.Then the relationship between the shot noise and ADC quantization noise in the QIS system is given.The image reconstruction of the initial image data without column noise is researched and the image restoration of a large number of binary data is realized.In addition,aiming at the noise characteristics of 3-D stacked parallel readout architecture,a block noise elimination algorithm based on convolutional neural network is designed.Finally,the validity of image reconstruction and noise elimination algorithm is proved by data experiments.According to the objective index,after the column noise elimination,SDCMV of 128 grayscale decreases from 2.6978 LSB to 0.9016 LSB,and SSIM increases from 0.8028 to 0.8641.After the block noise elimination,the PSNR and SSIM of 128 grayscale are increase from 26.4591 and 0.6733 to 34.6917 and 0.8036 respectively.