| Color constancy is the ability of humans to perceive the inherent color of an object’s surface,regardless of the influence of external light sources.However,existing photographic equipment does not possess this ability,leading to significant color cast in images taken under non-standard white light.Research on color constancy aims to give photographic equipment the ability to perceive color constancy,which has important practical applications in fields such as security surveillance,computational photography,and computer vision.The most advanced existing color constancy methods use deep learning technology to map RAW images and lighting labels,greatly improving the performance of color constancy.However,the field of color constancy faces problems such as difficulty in data collection and differences between multi-camera domains.This often results in insufficient data for existing deep learning methods,causing the model to learn false mapping relationships and limiting its expressive ability.In addition,existing methods also face problems such as high model parameters and computational complexity,which restrict the practical application of the model.To address the issue of insufficient data,we propose the Domain Adversarial Learning for Color Constancy(DALCC)network.The network consists of a domain adversarial learning branch and a feature reweighting module.In the domain adversarial learning branch,the camera domain classifier and feature extractor compete with each other in an adversarial way to encourage the appearance of domain-invariant features.At the same time,the lighting transformation module estimates the camera-related color space transformation matrix to perform color space conversion,solving the problem of inconsistent color spaces caused by domain-invariant features.The two work together to avoid the degradation of multi-camera domain models caused by domain differences in feature distribution,which allows our DALCC to benefit from multi-domain data.In addition,to better utilize multi-camera domain data,we propose a feature reweighting module to suppress non-primary light source regions,reducing the impact of misleading light source information on lighting estimation.The experiment shows that DALCC achieves the state-of-the-art performance on benchmark datasets by effectively utilizing multi-domain data.To address the problems of high model parameters and computational complexity of deep learning-based methods,we propose a color constancy network based on the deep separable Transformer.This network captures global information on the image to maximize the network’s receptive field,and uses the idea of deep separability in the self-attention mechanism to simplify calculations,reduce the computational complexity and model parameters of the network,and use point-wise convolution to improve the representation ability of the model between different channels.The experiment shows that the deep separable Transformer network achieves the same performance as the state-ofthe-art color constancy methods with the lowest network model complexity.Finally,from an industrial application perspective,we propose a color constancy system based on pure color scene classification to optimize the performance of extreme pure color scenes and provide solutions for deep learning model landing.This system uses pure color scene classification technology to separate non-pure color scene images from pure color scene images and uses deep learning algorithms to process pure color scene images to improve performance.Traditional algorithms are used to process non-pure color scenes to improve processing efficiency.In the pure color scene classification algorithm,we continuously explore the optimal classification features based on image lowlevel color and texture features,and study machine learning classification algorithms to obtain the best classification model.The system integrates functions such as image automation processing,scene classification,and color constancy algorithms,providing solutions for photography enthusiasts and practitioners. |
