Color Constancy Calculation Based On Convolutional Neural Network

In a range of illuminations,the human eye have the ability to perceive the surface color of objects in the scene.This perceptual characteristic that eliminates the color of the light source so that the human eye can accurately see the true color of the obj ect is called color constancy.Although the human visual system can easily identify objects under unknown light source and restore its surface color,computers do not have the characteristics of color constancy.In the field of computer vision,stable color feature is the basic condition of many computer vision tasks.Therefore,many scholars are committed to studying the problem of color constancy,so that the computer can also stably observe the actual color of the object surfaceBenefiting from the development of Convolutional Neural Network(CNN),substantial progress on color constancy has been made.This paper attempts to calculate color constancy by using CNN,and proposes algorithms based on strong supervision and weak supervision respectively,and has achieved satisfactory results.This article focuses on the following aspects(1)A color constancy method based on top-down semantic aggregation is proposed.High-level features of CNN structure contain semantic information while low-level features show local details.If both semantic and spatial information are taken into account,they would help achieve a more accurate illuminant estimation.Inspired by the pyramid model,a top-down network named TDCC is proposed.Firstly,this network successively propagates high-level information to low-level layers and obtain multi-scale feature maps with strong semantic information.Secondly,feature maps at each scale are utilized to estimate illuminant color via confidence-weighted pooling respectively.At last,these results are averaged to obtain the final illuminant estimation.Experimental results show that this method can identify objects with inherent colors and provide them with high confidence weights,which further improves the model's illuminant estimation accuracy and reduces the error of color estimate(2)A weakly supervised color constancy method based on generative adversarial learning is proposed.This method introduces Generative Adversarial Networks(GAN)to transform the color constancy problem into an image-to-image translation problem,that is,an image from an unknown light source is converted to an image under standard illuminant.This algorithm uses a loss function that combines color,content,and texture loss to learn the color of light source and recover image.Only two distinct datasets need to be prepared for network training one from the source datasets under unknown light source,and one composed of arbitrary images under standard illuminant that can be generally crawled from the internet.This article assumes that the normal images that can be obtained from the network are under standard lighting conditions.Experimental results show that this method has achieved satisfactory performance,it has smaller errors than most supervised algorithms.