Considering the lack of color constancy (CC) dataset in the field of microscopic images and the failure in achieving the expected effect through the cross-dataset training of the CC algorithm, this study creates a microscopic CC dataset using two steps: camera acquisition and simulation generation. Moreover, this study proposes a microscopic image CC algorithm based on an autoencoder. The algorithm uses an improved UNet structure autoencoder for semi-supervised training and simultaneously introduces a new composite-loss function to optimize network parameters, thereby obtaining an accurate restored image color. Experimental results show that the image resolution trained using the algorithm is higher than traditional autoencoders, and the angle error estimates in the NUS-8, RECommended, and self-built microscope CC datasets are smaller.