Research On Color Constancy Calculation Based On Deep Learning

In the field of computer vision,color is a crucial feature that is widely used in various tasks such as autonomous driving and security monitoring.However,since the color of the object will be affected by the scene illumination,the color of the object presents an obvious color cast.In contrast,the human visual system has the perceptual characteristics of color constancy,which can help us to recognize the true color of objects,that is,to recognize consistent colors under different illumination conditions.The purpose of color constancy research is to simulate the visual characteristics of the human eye to automatically remove scene illumination,obtain accurate colors of objects,and provide stable color features for other computer vision tasks.This thesis discusses the problem of color constancy calculation under the condition of single illumination,and adopts deep learning technology to solve the shortcomings of existing color constancy calculation methods.The main work of this thesis is as follows:(1)Research on traditional color constancy calculation.This thesis compares and analyzes a variety of statistical-based color constancy algorithms,and makes an in-depth and detailed analysis of the basic ideas,basic principles and correction effects of each algorithm.By comparing the advantages and disadvantages of several algorithms,it provides a basis for the color constancy calculation method based on deep learning proposed later.(2)In view of the inaccuracy and instability of illumination estimation in the existing color constancy calculation methods,a color constancy calculation method of Conv Ne Xt network with fusion confidence is proposed.First,Conv Ne Xt is used as the backbone network,and the confidence weighting method is added to the backbone network.The confidence weighting method calculates the confidence of each image patch and determines the adjustment weight according to the region of the image patch,so as to increase the network’s attention to key features and reduce the attention to non-important information.Conv Ne Xt network structure can effectively avoid the loss of feature information and more fully utilize the information in the channel for illumination estimation.This enables the network to perceive illumination information more accurately,thus significantly improving the accuracy of the illumination estimation algorithm and showing stronger robustness when dealing with complex environments.(3)The color constancy calculation method based on convolutional neural network has some limitations.Its receptive field is small and fixed,which leads to ignoring the feature information outside the receptive field in the image,and it is difficult to obtain the global information of the image.To this end,a color constancy calculation method based on hierarchical network is proposed.This method improves the Swim Transformer network and uses a hierarchical network architecture to extract multi-scale feature maps,so that the information of the entire image can be obtained more comprehensively and effectively.This multi-scale feature extraction method makes the method more robust and adaptable when processing images of different sizes and illumination.At the same time,the method uses a moving window to calculate self-attention and fully considers the relationship between pixels,thereby more accurately estimating the scene illumination,and improving the efficiency and accuracy of the algorithm on the premise of reducing the computational complexity.