【摘要】 消除由于光照等条件变化而对图像颜色值产生的影响是进行彩色图像的识别和分析的关键 .该文从表面反射的有限维线性模型出发 ,在理想情况下导出了颜色值从非标准光照到标准光照的线性转换关系 .实际上 ,由于成像系统本身采用了一些处理技术 ,使得转换关系应为非线性 .文中提出了新的监督颜色校正方法 ,通过在图像摄取环境中放置监督色板 ,并借鉴监督学习的思想 ,分别采用线性回归、多项式回归和BP神经网络三种方法来求解转换关系 ,从而对颜色值进行校正 .试验结果表明 ,三种方法都达到了较好的校正效果 ,且应用条件几乎不受限制 .相比之下 ,多项式方法优于其它两种方法更多还原

【Abstract】 In this paper, a linear transformation relationship between color value taken in a canonical illuminant environment and one in an unknown illuminant environment is derived from the finite dimensional linear model. Actually, the transformation relationship must be non linear taking into account some processing techniques applied by the color camera system itself. Supervised color correction, which uses color patch to calculate and correct illumination changes, is used widely because of its simplicity and efficiency. We propose three new supervised color correction methods based on the ideas of supervised learning. Firstly, the supervised color patch containing 33 color chips in different chroma and low saturation is designed. Then those color values of color chips under canonical illumination and unknown illumination are recorded in order to obtain the set of learning samples. Finally, the transformation relationship is determined respectively by linear recursion analysis, polynomial recursion analysis, and BP neural network mapping according to the set of learning samples. These methods differ from the previous ones in that they do not need to know the spectral reflectance functions of color patch and the spectral sensitivity and zero coordinate of the camera. To test the performance of these methods, we did some experiments on 15 color objects with different color and different quality. Though the color difference of the color object in different ambient light is big, it obviously become smaller after color correcting with these methods. The results show that the three methods all have good color constancy, especially, the polynomial recursion analysis is better than the other two.更多还原