Research On Flat-field Scheme Of Color Image Sensor

At present,the color image sensor integration process is very mature and is widely used in various fields such as aerospace,remote sensing imaging,and industrial monitoring.Color CMOS image sensors are widely used due to their advantages of high integration,low power consumption,and simple driving.However,the non-uniform noise generated by the inconsistency of the pixel response of the color CMOS image sensor will lead to the nonuniformity of the output image,which will seriously affect the imaging quality.Secondly,the crosstalk phenomenon between pixels of different colors will cause the pixel photon response curve of the image sensor to show nonlinear characteristics,reduce the signal-to-noise ratio of image data,and cause the loss of image information.Existing research can complete the flat-field correction of grayscale image sensors,but for color image sensors with heterochromatic pixel crosstalk,traditional correction methods have the disadvantages of insufficient accuracy and inconspicuous correction effects.In this paper,the pixel response model is established,and the offset correction and gain correction are performed after compensating the amount of crosstalk between pixels to obtain the ideal output value of the pixel.Using the pixel measured gray value data and pixel ideal gray value data as the learning set for neural network training,the trained neural network is used to perform flatfield correction on the color CMOS image sensor.In order to realize the acquisition and correction algorithm of pixel grayscale data of image sensor,a set of color image sensor test system is designed and built.The entire system structure is divided into two parts: the upper computer system and the lower computer system.The upper computer system is composed of data acquisition software,and the lower computer system mainly includes data acquisition circuits,light sources,and light intensity sensors.Through the interactive control of the upper computer and the lower computer,the setting of exposure,the collection of image grayscale data and related calculations are completed,providing a hardware basis for subsequent image sensor calibration.On the basis of photon conversion theory,a neural network flat-field correction method based on crosstalk compensation of neighboring pixels is proposed.This method compensates the amount of crosstalk between pixels of different colors,and corrects the gain coefficient and offset coefficient of each pixel.The ideal value of the flat field of the pixel is obtained,which solves the problem of large errors caused by the same correction parameters of each pixel,and improves the accuracy of the flat field correction.According to the Bayer array arrangement rule of the color image sensor,a neural network is built for training,the original image grayscale data is used as the input of the neural network,and the ideal gray value after processing is used as the output value,and the neural network is built for experiments.After the preliminary construction of the neural network is completed,the structure and parameters of the network model are optimized,the influence of each structure and parameters on the network is characterized,and the optimal network structure and parameters are selected.After optimization,the verification data set experiment is carried out,the loss value of the neural network is small,and the convergence is obvious,which confirms the effectiveness and feasibility of the neural network.The real shot image is verified.From the analysis of traditional objective image evaluation indicators,after correction by this method,the peak signal-to-noise ratio is 55.928 d B,the root mean square error is 0.183,and the flatness is 0.0111.Compared with the image before correction,the peak signal-to-noise ratio The noise ratio is increased by 9.03 d B,the root mean square error is reduced by 0.969,and the flatness is reduced by 0.0695.According to the principle that the monochromatic light uniform exposure image after flat-field correction should contain lower contrast,Fourier expansion is performed on the grayscale data of the image before and after correction,and the information in the image space domain is converted into information in the frequency domain.Comparing the image spectrograms before and after correction,we can see that,After correction,the high-frequency information of the image is reduced,and the image uniformity is better.The flat-field correction scheme proposed in this paper applies neural network for flatfield processing on the basis of pixel crosstalk compensation,and proposes a new method to evaluate the effect of flat-field correction.The experimental results have achieved good results in subjective observation and experimental data evaluation index verification,indicating that this scheme has a good flat-field correction effect on color CMOS image sensors.