Research On Image Color Constancy Algorithm Based On Deep Learning

Color constancy means that when the color of the light source shining on the surface of an object changes,human's perception of the color of the surface of the object remains stable.Color constancy algorithms can be implemented by estimating the illuminant color and using it to transform the original image.Color constancy algorithms are more and more important in various image applications and computer vision algorithms.Correcting images using color constancy algorithms can not only make them more aesthetic,but also enhance them for further computer vision algorithms.A color constancy algorithm based on fully convolutional network(FCN)has been proposed in this study,which features an Adding Pooling layer,so that it can take the complete image as input and learn its semantic information fully.Visual attention modules are designed to improve the pormance of the network.Finally,the algorithm is optimized for mobile application and deployed on mobile terminals.The main work and innovations of this thesis list as follows:(1)Designing a color constancy algorithm based on FCN with Adding Pooling layer.Existing color constancy algorithms based on CNN often use local image patches as input to obtain local estimates,then integrate all local estimates into global estimates.Local image patches contain little semantic information which account for invalid inference.Meanwhile,most networks include fully connected layers which require fixed size input,leading to the cropping or deformation of test images,and furthur resulting in the loss of semantic information.The incompleteness of test images will cause the reduction of the accuracy of prediction results.In this study,FCN is used to implement the color constancy algorithm.Full image is used as input.Fully connected layer is removed which enable the input to be any size.The Adding Pooling layer is proposed,whose affect is to adaptively weight tlocal illumination estimates and add them into global illumination estimates,which effectively solves the problem of the lack of semantic information in local image patches and distortion of test images,the result is significantly improved.Results on the Color Checker Dataset and the NUS 8-Camera Dataset show that the algorithm proposed in this study is significantly better than the current mainstream algorithms.(2)Designing attention modules for the color constancy deep learning network proposed in this study.A lightweight channel attention module and a lightweight spatial attention module are designed,and two mixed attention modules are designed based on them,which are combined in series.The performance of the model is obviously improved while the size and inference duration are not obviously increased.(3)Optimizing the color constancy algorithm proposed in this study for mobile application and deploying it on mobile terminals,verifying the effectiveness of the algorithm and the feasibility of the optimization scheme.Data augmentation and model compression are used to optimize the model.The device correlation of the color constancy datasets is analyzed and removed by data augmentation,which makes the trained model can be applied to any image taken by any device and have stronger generalization ability;Model compression is performed by deleting non-test nodes and parameters and weight quanting,results show that the model size can be compressed from 141MB to 6.66MB without affecting the performance of the algorithm,which significantly reduces the resource requirements of the model for mobile devices;The designed algorithm is used in Android application to carry out real-time color constancy for the photos selected or taken by users,which achieves an excellent correction effect and a fluid user experience.