Objective method with deep learning to calibrate the spatial color resolution of a color charge-coupled device

An objective method to calibrate the spatial color resolution (SCR) of a color charge-coupled device (CCD) is provided. The experimental prototype contains a target generation system, a test platform, and an analytical processing system. The target generation system creates colors with an adjustable light source and two complementary resolution panels (each one has target bars). The test platform aims at creating images through adjusting parameters of a light source of the target generation system. And the analytical processing system is used to process images to evaluate the SCR of a color CCD. We focus on the third module and utilize the minimum detectable color difference (MDCD) and the minimum resolvable color difference (MRCD) to evaluate the SCR. In the process of data collecting, we first set the two channels (one is foreground channel and the other is background channel) of a generation system the same color and then gradually change wavelength of the foreground channel until the foreground image is slightly visible. As it is difficult to let the ratio of detectable pixels of target bars just meet the requirements of the MDCD and the MRCD by only adjusting the wavelength, we adopt the general regression neural network to estimate the two indicators and the maximum estimation error of which are within 6%. In order to deal with more complex scenarios with brightness and saturation change, an image augmentation network (a modified generative adversarial network) is applied to generate synthetic images, which cannot be easily captured by our prototype due to limits of the light source. The experimental results show that the estimation error of MDCD and MRCD is decreased to almost 1%. The method is a human-eye independent way and performs well in selecting the right kind of a color CCD, which guarantees the reliability and security of the visual detection and recognition system.