Algorithm Research And System Implementation Of Color Constancy Based On Transfer Learning

The research of Color Constancy is the cross research direction of computer vision and human perception.Its purpose is to make the computer have the robustness ability to the objects’ color and to correct the image color cast caused by scene illumination,like the human visual system.With the popularity of digital images on the Internet and the rise of computer vision,more and more scientists,photographers,camera equipment manufacturers,and even aerospace need to carry out white balance processing on images in order to ensure the authenticity of the image,so as to facilitate the subsequent analysis and research of scenes or objects;In addition,the latest research in 2021 shows that white balance will also affect the quality of various downstream visual tasks,such as semantic segmentation and classification.The mainstream approach for Color Constancy tasks is to estimate the color of illumination from the RAW image and then correct it by Von Kries model.In recent years,the application of deep learning methods to singlecamera data has made remarkable progress.However,due to the expensive data collection for color constancy research,these models still have a seriously insufficient data problem,resulting in a shallow model capacity;in addition,due to the differences in the spectral sensitivity of camera sensors,the data captured by different cameras have obvious chromatic aberration,which will lead to more difficult to fit the model.In this paper,in order to alleviate this problem,we propose a Transfer Learning Color Constancy(TLCC)method,which can use RAW data of multiple cameras and a large number of unlabeled sRGB data to support training,avoid sensor-domain gap differences and obtain scene information from sRGB data.Specifically,the color constancy algorithm based on transfer learning is composed of Statistical Estimation Scheme(SE-Scheme)and Color-Guided Adaptation Branch(CGA-Branch).SE-Scheme builds a statistical perspective,maps camera-related lighting tags to camera-agnostic forms,and generates pseudo labels for sRGB data,which avoids the differences among multiple cameras from the tag perspective and greatly expands the data of joint training.Based on the transfer learning method,CGA-Branch extracts the unique color information of the image and adaptively adjusts the characteristics of the feature by controlling the scale and shift parameters in the normalization module,which further promotes efficient transfer learning from sRGB data to RAW data.The experimental results show that TLCC overcomes data limitations and model degradation,and achieves overall model improvement when the model stacks up to 2 layers of basicmodules and mixes training data.Finally,based on the research results of the TLCC,this paper designs a set of Color Constancy Domain Systems,which includes an Automatic Labeling Subsystem to automatically process collection data,filter unqualified images and establish a data set.It also includes the Color Constancy Subsystem,which provides the ability to perform image pre/post-processing,automatic white balance,and color correction on the PC side,so that the PC side can horizontally compare the final correction effects of various scientific research algorithms actually applied to the camera pipeline.In addition,in order to actively promote the open-source work in the field of Color Constancy,this paper designs the first open-source code library,Anole,which contains the most representative baseline models in the past 5 years,and also has rapid training,standardized testing,model secondary development,etc.,which can provide a good entry code for researchers and application personmnel,and lay a foundation for the later unified/standardized model test work.