Research On Color Constancy With Machine Learning

In a range of illuminations,human have the ability to perceive the true color of objects in the scene.This ability to diminish the effect of illumination,and make human perceive the true color of the object's surface is called color constancy.Color constancy is human's instinctive ability in the long period of evolution.Human's excellent color constancy plays an important role for human to learn the world.In the field of computer vision,the correct color feature is the basic condition of many computer vision tasks,but the camera is easily affected by the illumination of the scene when the camera generating the image.The color of the same object in the image will be different due to illumination,this lead to color feature instability,so the color constancy is an important research field.This paper attempts to use machine learning technology to solve the difference of the image under different illuminations and restoring true color.We do research on following aspects.Respect to the shortcomings of traditional diagonal correction models,ground truth color restoration algorithm based on full matrix correction is an effective color constancy recovery algorithm.We analyze the existing problems and shortcomings in this algorithm,and propose ground truth color restoration algorithm based on image content.First,classifying all image pixels by color,the algorithm counting up the proportion of various colors in the graph,weighted the optimization formula for finding the full matrix,which makes the algorithm overcome the instability in a way,and improve the accuracy of color correction.But the algorithm is only used for the datasets with color checker,the use of algorithm is restricted.And then this paper implemented the existing color constancy algorithm based on neural network is introduced.Although the algorithm can perform quite good,the network is lack of interpretability.Based on this,we use visual method to get the features learned by the network,and proposed a new visualization method according to the characteristics of color constancy neural network.In this method,the input image is scanned in a small stride block by block,and the obtained blocks are send into the neural network.The activation values of the corresponding neurons are obtained by forward propagate,and all the activated values are arranged according to the position,reforming the activation matrix and then converted into a thermodynamic diagram,finally the visual view is obtained.The algorithm overcomes the shortcoming of deconvolution visualization method in color constancy network,and new method can get the clearly response of features in the network.According to the visualization tools,the work mode of the network is analyzed and explained.The meaning of each step in the network is explained,and how the content of the image affects the final prediction results are revealed.