Research On CMOS Non-uniformity Correction Method

Complementary metal oxide semiconductor(CMOS)image sensor has been superior to charge coupled device(CCD)image sensor in frame rate,power consumption,integration,cost-effective and other aspects.In recent years,CMOS image sensor has been continuously improved in technology and structure design,making it have significant improvement in readout noise,dark current,dynamic range,quantum efficiency and other disadvantages.Recently,CMOS has been widely used in various high and low-end fields.Unfortunately,due to the readout structure,CMOS has a more significant nonuniform noise than CCD.Based on CMOS observations,the Yunnan Observatory has found that the time-varying noise can be effectively suppressed in the reconstruction process,while the relatively stable spatial noise(non-uniform noise)is further strengthened,resulting in the final reconstruction result being seriously affected.Therefore,in order to further improve the quality of high-resolution images,the correction of CMOS nonuniformity becomes an unavoidable problem.Nonuniformity correction methods can be divided into calibration-based method and scene-based method.Due to the influence of temperature drift and time drift,calibrationbased methods will face the problem of calibration parameter failure.Therefore,in order to maintain the validity of the calibration parameters,it is necessary to calibrate the parameters repeatedly.The flexibility of this method is seriously affected.In order to overcome the lack of flexibility of the calibration method,the scene-based method has become a research hotspot.According to different application requirements,three types of scene-based nonuniformity correction methods are proposed.After analyzing the causes of non-uniform noise and the characteristics of noise image on the related wavelet components,this dissertation presents a non-uniformity correction method based on the wavelet principal component analysis.The method considers that the presence of nonuniform noise causes the change of DC component in the correlated wavelet component.Therefore,this method realizes the readjustment of DC component mainly through dimension reduction and noise reduction,and then realizes the correction of non-uniformity.The fast calculation speed of this method is more suitable for application scenarios where the correction efficiency is required.This dissertation presents a non-uniformity correction method for low rank sparse variations after analyzing the structure and distribution of non-uniform noise.This method realizes the correction of non-uniformity by solving an optimization program that combines prior information such as non-uniformity noise low-rank characteristics,distribution Gaussian characteristics,target image variational sparse characteristics,and fidelity characteristics.At the same time,in order to simplify the selection of the regularization coefficients,this method introduces an adaptive coefficient adjustment strategy based on low frequency information,which can dynamically balance the intensity of fidelity with the intensity of image smoothing,and can reduce the adjustment difficulty of the regularization coefficients to a certain extent.This method has good correction performance,but the correction process is relatively time-consuming and is suitable for application scenarios where the correction performance is required but not sensitive to the correction efficiency.This dissertation combines the excellent feature expression ability of residual network with the anisotropic characteristics of image,and presents a method for nonuniformity correction of full-variation residual network.The network removes the nonuniform noise image from the input image by continuously learning the mapping relationship between the measured sample data and the label data obtained by the segmented calibration method.Although this method requires specialized equipment to generate training data and a long previous network training process,the fully trained network not only has the same non-uniformity correction efficiency as the filter method,but also has the same correction effect as the optimization method.This method is suitable for application scenarios where both correction performance and correction efficiency are required.