Research On Image Sensor Correction Algorithm Based On Neural Network

Image sensors are widely used in electronics,aviation,aerospace,medical and military.In an ideal state,the ideal photon transfer curve of image sensor is a linear function,meanwhile,the gray value of different pixels in a image sensor should be the same.However,the actual photon transfer curve is non-linear,and the dynamic range of the image sensor is different from the ideal state,meanwhile,the actual gray value of the pixel is different from the ideal value,which causes the contrast and uniformity of the captured gray image is insufficient..Therefore,the image sensor needs to be corrected.Although the current methods could correction the non-uniformity of the image sensor,the accuracy is not high enough.This paper proposes an image sensor correction algorithm based on neural network to correct CCD or CMOS image sensors.The neural network can generate complex correction coefficients to implement complex algorithms so that it could obtain better correction results.The image sensor test system is built to implement the correction algorithm.The system includes: upper computer and lower computer,the lower computer is data acquisition circuit,the upper computer is an operation software,and the upper computer communicate with the lower computer by the Camera-Link interface.Image acquisition,photon transfer curve drawing and related calculations can be realized by the image sensor test system.An algorithm was proposed to find the optimal exposure step automatically for the reason to obtain more effective image data.Gray images are captured in 100 exposure conditions with the optimal exposure step as a time interval.Then we will analysis the feature of the gray image,after that,the picture are preprocessed to repair the damaged rows or columns of the CCD / CMOS image sensor.The pixels are divided into 10 categories before training the neural network in order to improve the accuracy of the algorithm,and a neural network is trained for each category of pixel.The classified data is the input of each neural network,and regard the corresponding ideal gray value as the output data.After the experiment,we will optimize the model structure and parameters,and the best parameters and structure of the network are taken as the final network structure.The optimized experimental results show that the model can confirming the feasibility and effectiveness of the proposed model.After the neural network training,the algorithm is tested by capture pictures.Comparing the images before correction to after correction,it can be seen that:(1)the dynamic range of the corrected PTC curve becomes 0-4095 DN,which shows that the algorithm improve the dynamic range of the image sensor;(2)the grayscale of the image sensor are uniform,and the inconsistency and non-linearity of the image sensor are improved;(3)the defected pixels are well processed,and the uniformity of the corrected image is better;(4)The sharpness and contrast of the pictures after correction are improved.The analysis of the corrected image shows that the variance of the corrected image is bigger.The correction algorithm increases the range of the gray distribution and the dynamic range of the image sensor.The corrected image shows high contrast and clear borders,the picture is good performance in details,and the definition of the image has been greatly improved.Experimental results show that the algorithm greatly improves the performance of the image sensor,so it can be seen that the algorithm we proposed has good practicability and effectiveness.