电子测量技术ELECTRONICMEASUREMENTTECHNOLOGY第45卷第23期2022年12月DOI:10.19651/j.cnki.emt.2209925综合利用肤色模型和感知损失的偏色去除*桂冰洁贾振堂(上海电力大学电子与信息工程学院上海201306)摘要:由于现有的颜色恒常性算法在非均匀光照、场景复杂的情况下对于偏色去除表现不好,本文提出了一种综合利用肤色模型和感知损失去除偏色的深度学习算法。该算法综合利用了肤色模型和感知损失,在计算过程中可以识别且重点关注肤色信息,更注重对图像语义的理解,而不是简单的进行像素间的计算。同时,将肤色模型与注意力机制结合,更突出了肤色区域的作用。实验结果表明,本文提出的颜色恒常计算方法能在语义层面较为准确地消除单一光照和多光照场景下图像的偏色,该算法与其他算法相比,能获得更良好的效果。关键词:颜色恒常性;图像处理;图像恢复;感知损失中图分类号:TP391文献标识码:A国家标准学科分类代码:510.4050ColorcastremovalusingskincolormodelandperceptuallossGuiBingjieJiaZhentang(CollegeofElectronicalandInformationEngineering,ShanghaiUniversityofElectricPower,Shanghai201306,China)Abstract:Sincetheexistingcolorconstancyalgorithmsdonotperformwellforcolorcastremovalinthecaseofnon-uniformilluminationandcomplexscenes,thispaperproposesadeeplearningalgorithmthatcomprehensivelyusestheskincolormodelandperceptuallosstoremovecolorcasts.Thealgorithmintegratestheskincolormodelandperceptualloss,sothatcanrecognizeandfocusonskincolorinformationinthecalculationprocess,andpaymoreattentiontotheunderstandingofimagesemantics,ratherthansimplecalculationbetweenpixels.Atthesametime,theskincolormodeliscombinedwiththeattentionmechanism,whichhighlightstheroleoftheskincolorarea.Theexperimentalresultsshowthatthecolorconstancycalculationmethodproposedinthispapercanaccuratelyeliminatethecolorcastofimagesinsingle-illuminationandmulti-illuminationscenesatthesemanticlevel.Comparedwithotheralgorithms,thisalgorithmcanachievebetterresults.Keywords:colorconstancy;imageprocessing;imagerestoration;perceivedloss收稿日期:2022-05-12*基金项目:国家自然科学基金(62105196)项目资助0引言人眼依靠感光细胞观察事物,光线反射到人类双眼经由一系列处理,在大脑中形成图像,以识别物体、认...
