光谱成像可以记录拍摄场景的光谱信息, 从而实现颜色的高保真复现, 图像的光谱信息通常同时混合了物体的光谱信息和光源信息, 所以现有的光谱成像颜色还原需要通过提前放置的标定板或已知光谱特性的物体获取拍摄环境的光源信息, 再进行图像的颜色恒常性校正。 但光谱相机在实际使用过程中, 通常难以满足上述条件, 从而对图像高保真颜色还原提出挑战。 针对此问题, 提出一种光谱成像的图像颜色恒常性计算方法, 将光谱成像系统获得光谱数据转换到XYZ颜色空间, 并将图像分区域进行统计得到图像的统计点, 通过大量常见光源的分布规律, 给统计点施加位置权重和色温权重, 同时设置亮度权重去除图像中过暗和过饱和的统计点, 利用加权平均获得环境光的色度参数, 根据需求将图像的XYZ颜色空间数据转换到RGB颜色空间, 根据环境光的色度参数计算图像不同通道的增益, 从而完成光谱图像的颜色恒常性计算。 为了验证所提出算法的有效性, 对140张光谱图像进行处理, 计算算法得到的环境光色度参数和真实光源色度之间的再现角误差, 对校正结果进行评价, 结果表明提出的算法明显优于光谱图像不做处理和灰度世界法。 为进一步分析算法校正结果与人眼感知之间的联系, 设计颜色心理物理学实验中的分度实验, 18名视觉正常的观察者参与实验, 所有观察者对算法校正结果打分的平均值介于良好和优秀之间, 可以满足实际使用需求； 沿着平均色温线的方向, 观察者所允许的光源色度差异较其它方向稍大； 在图像中包含大面积记忆色时, 当光源色度差异向使记忆色饱和度增大的方向偏移时, 此时观察者所容忍的色度差异变大。 由客观和心理物理学实验结果可知, 所提出的算法可以在拍摄光源未知, 且没有标板的情况下, 较好的对光谱成像中的图像颜色恒常性进行处理, 为光谱图像的颜色高保真再现打下基础。

Spectral imaging can record the spectral information of the scene to achieve high-fidelity color reproduction. The spectral image is usually mixed with the spectral information of the object and the light source information at the same time, so the existing color reproduction methods of spectral imaging need to obtain the light source information of the environment through the calibration target placed in advance or the object with known spectral characteristics, and then correct the color constancy of the image. However, it is usually difficult to meet the above conditions in the actual use of spectral cameras, which brings challenges for image high-fidelity color reproduction. A calculation method of image color constancy for spectral imaging was proposed. The spectral data obtained by the spectral imaging system were converted into the XYZ color space, and the image was divided into regions for statistics to get the statistical points of the image. According to the distribution rule of many common light sources, position weights and color temperature weights were applied to statistical points, and luminance weights were set to remove over-dark and over-saturated statistical points in the image. The chromaticity parameters of ambient light were obtained by weighted average. Moreover, the XYZ color space data of the image were converted to RGB color space according to the application requirements. The gains of different channels of the image were calculated according to the chromaticity parameters of ambient light to complete the spectral image’s color constancy calculation. In order to verify the effectiveness of the proposed algorithm, 140 spectral images were processed, and the reproduction angle error between the chromaticity of the ambient light obtained by the algorithm and the chromaticity of the real light source was calculated. The correction results showed that the proposed algorithm was better than the spectral images without processing and the Gray-world method. In order to further analyze the relationship between the algorithm correction results and human perception, a color psychophysics experiment was designed, 18 observers with normal vision participated in the experiment. The average score of all observers for the algorithm correction results was between good and excellent, which can meet the actual use needs. Along the direction of the average color temperature line, the chromaticity difference of the light source accepted by the observer was slightly larger than in other directions. When the image contained a large area of memory color, and the chromaticity of the light source shifted in a direction that increased the saturation of the memory color, the chromaticity difference accepted by the observer became larger. The results of objective and psychophysical experiments show that the proposed algorithm can deal with the image color constancy in spectral imaging well when the light source is unknown, and there is no calibration target, which lays a foundation for high-fidelity color reproduction of spectral images.