Research On Four-band Image Color Correction Algorithm Based On Lasso Regression And Extreme Learning Machine Algorithm

CCD or CMOS sensors in traditional image acquisition equipment can sense visible light and near-infrared light.Thus,in the process of shooting in the daytime,the infrared ray will interfere with the image sensor and make the image red as a whole.In this thesis,we call this kind of color cast image mixed with near-infrared light in the visible light as four-band images.In order to achieve 24-hour shooting,the traditional monitoring system will add the infrared cut-off filter(IR-CUT)which can be mechanically switched in the imaging system.The built-in IR cut-off filter can automatically switch according to the external light intensity.The image acquisition equipment can turn on the infrared cut-off filter in the daytime to block the infrared light and turn off the filter at night to make the use of infrared lighting for imaging at night.Because of the damage,failure,and high cost of replacing the devices,this thesis considers discarding the mechanical switch system and use algorithms to remove the interference of infrared.In this thesis,on the basis of previous research on the four-band images,a four-band image color cast correction algorithm based on Lasso regression is proposed.The color correction method based on the supervision is to place standard color cards with rich colors in the light source of the scene,take the chromaticity values of these color cards in standard light source and non-standard light source as learning samples,use the color cards to establish the regression equation relationship between the gray values of the three-band image color block and four-band image color block in R,G,and B channels,and use the least square method to fit the correction matrix.And then,the correction matrix is used to correct the four-band images under different light source color temperatures and different scenes.Lasso polynomial regression fitting algorithm based on its own advantages uses the absolute value function of the model coefficients as the penalty term to compress the coefficients.It can simplify the model while retaining the influential characteristic variables in the polynomial,effectively improve the data fitting accuracy on the premise of ensuring the complexity of the model,thus enhancing the explanatory power of the model,and has strong applicability.On the other hand,due to the influence of various factors such as the image acquisition equipment itself,the color temperatures of the environment light sources,and the different reflectance of the object to light,the conversion relationship of the image pixel values are greatly affected by the nonlinear factors.Polynomial regression algorithm cannot meet the high requirements of nonlinear processing of color cast images at higher color temperatures due to its own limitations.To improve its correction accuracy,this thesis introduces the method of polynomial regression Extreme Learning Machine to improve its nonlinear mapping ability.Like polynomial regression algorithm,extreme learning machine correction algorithm constructs the mapping relationship between a large number of three-band images and four-band images based on training and achieves the purpose of overall correction of four-band images by obtaining the correction weight matrix of the three-band image and the four-band image pixels mapping.The experimental results show that the color cast correction algorithm based on the learning machine has a good correction effect on the outdoor four-band images with green plants seriously affected by infrared crosstalk,which effectively solves the problems that the correction effects of the outdoor four-band images with green plants are not ideal.Moreover,the model constructed by the algorithm has strong generalization ability and fast image correction speed,which lays the foundation for the practicability of the four-band image correction algorithm in the future.