Abstract: To solve the problems of color overflow and insufficient detail preservatio in gray-scale image coloring using generative adversarial network, a pollen gray-scale image coloring method based on the cycle-consistent generative adversarial networks (CycleGAN) is proposed in this paper. This method colors grayscale images using CycleGAN from the perspective of unsupervised learning. To solve the problems of discontinuous coloring and non delicate coloring in pollen gray-scale images, non-local modules were introduced to effectively obtain global information representation of the image. In addition, a self-attentive mechanism was introduced to help the network more accurately determine the spatial position relationship between pixels, thereby enhancing the learning ability of the coloring model. Experiment results show that the peak signal-to-noise ratio, structural similarity, and mean opinion score obtained by this method are 28.673, 0.956, and 4.567, respectively. The proposed gray-scale coloring method produces better color image quality on the test image set. This method not only effectively solves the problems of color overflow and discontinuous coloring, but also enriches the image details.