基于BP神经网络的数码相机特性化

由于数码相机的颜色空间是依赖于设备的,对于一个具体的数码相机,其光谱响应与设备独立的CIE标准观察者颜色匹配函数是一个非线性关系,因此不能真实复制场景的颜色。特性化彩色图像设备是提高图像的颜色复制质量的一个重要方法。介绍一种基于BP神经网络数码相机特性化方法。采用Munsell颜色系统作为目标色,大样本训练空间。测试了不同的网络结构和样本空间分布。训练样本平均色差为1.75CMC(1∶1)色差单位,测试样本为2.16。该方法在数码相机颜色测量、光谱重建等领域有广泛的应用前景。

Characterization of still camera BP neural networks-based

The color space of camera is device-dependent. Furthermore, their spectral respond do not directly correspond to the device-independent color matching function based on the CIE standard colorimetric observer. The camera could not faithful reproduction the color of scene. The color device characterization is such an importance device-independent way to improve the color accurate on captured images. The BP neural networks are employed to achieve camera characterization. Munsell book of color is adopted for the color target. A larger samples space is used to train networks. The difference architecture of networks and space of samples are tested. The average color difference of training samples is 1.75ΔE units in CMC(1∶1), and testing samples is 2.16 when a 6-20-20-20-3 network architecture is used. The method can be widely applied for camera color measurement and spectral reconstruction etc.