Efficient spectral reconstruction using a trichromatic camera via sample optimization

Training-based multispectral reconstruction can effectively recover spectral reflectance of captured objects using a trichromatic camera. However, existing methods are based on synthesized data, and the sizes of training sample set (e.g., multispectral images, reflectance targets) are usually large. In this paper, we present a spectral reconstruction approach using real measured data. To improve the efficiency and accuracy of spectral reconstruction, we propose a volume maximization method for sample optimization without any prior knowledge of light and cameras. We use heuristic global search algorithms to optimize samples and give an efficient spectral reconstruction method which is suitable for sparse sampling. Experimental results show that the proposed sample selection method outperforms other existing methods in terms of both spectral and colorimetric reconstruction errors. Moreover, the proposed reconstruction method achieves higher efficiency and accuracy due to lower sample redundancy.