<p>Image colorization is a hot topic in the research of image processing and generation techniques. It plays an important role in restoring old photos, colorization medical images, and colorization anime manuscripts. In recent years, with the rapid development of deep learning, some colorization methods based on deep learning have been proposed, and satisfactory colorization results have been achieved. However, these colorization methods generally use unguided feature learning to achieve color prediction, it cannot predict, select, and intervene in image colors based on the subjective ideas of the creator. At the same time, the generated colorization results may have unreasonable phenomena such as color concentration, saturation and brightness distribution. Therefore, we propose a cross-domain color mapping from exemplar anime image colorization method, which is used for coloring a given anime image with a reference style. This method involves a supervised colorization network, which is divided into two structures: a domain semantic matching sub-network and a colorization sub-network. The domain semantic matching sub-network maps the pixels of the input grayscale image domain and the reference color image domain to an intermediate domain by calculating the similarity of the correlation matrix, thereby establishing a corresponding semantic relation and obtaining a coarse colorization map. We also employed an end-to-end CNN encoder and decoder network to further extract the feature information of the matched image colors, and used the Lab color space to select, propagate, and predict the color distribution features of the image. Experimental results have demonstrated that the anime coloring images generated by our network have high restoration similarity compared to the original reference images, reasonable coloration in the coloring area, and this colorization method solves the problem of prior color style transfer in anime images.</p>