颜色恒常性是实现识别、分割和三维物体重建等视觉任务的重要前提。为使计算机视觉系统具有颜色恒常性感知功能,提出多通道特征置信度加权网络,在减少网络层数和模型参数的同时充分提取图像中的特征;通过多通道置信度加权方法利用每个通道中可以为光源估计提供更多信息的特征准确估计出全局场景光源。在基于重处理的ColorChecker和NUS-8数据集上的实验结果表明,本文算法通过对特征从多通道进行置信度加权,在各项评价指标上均优于目前的颜色恒常性算法,提高了算法的精确性和稳健性,可应用于需要进行色彩校正的计算机视觉任务。

Color constancy is an important prerequisite for visual tasks such as recognition, segmentation, and three-dimensional object reconstruction. We proposed a multi-channel feature-confidence-weighted network to enable the computer vision systems to perceive color constancy. As a result, the network could fully extract the features in the images while reducing the number of network layers and model parameters. The multi-channel confidence-weighted method employed the features in each channel that could provide more information for light source estimation to accurately estimate the light source in the global scene. Experimental results on the reprocessed ColorChecker and NUS-8 datasets show that the proposed algorithm, which weights the confidence of features in multiple channels, outperforms its counterparts in terms of all evaluation indexes and thus has higher accuracy and robustness. As such, this algorithm can be applied to the tasks of computer vision requiring color correction.