Image style transfer based on improved detail retaining

In this paper, the detail-retaining plug-in (DRP) connects the shallow features output from the encoder, and makes the shallow features participate in the attention computation through the divide-and-conquer mechanism, the slicer firstly slices the shallow features into a number of sub-features losslessly, and then feeds them into the style transformation module, so that the shallow features attend in the attention computing the style transformation with a small amount of computation cost; In the style transformation module, the mean and standard deviation of the content features are adjusted to align with the mean and standard deviation of the attention-weighted stylized features, and the stylized features are output; then the shallow stylized features are reconstructed back to the original dimensions by the splicer; finally, the reconstructed shallow stylized features are sent to the decoder along with the deeper stylized features, and the decoded outputs are obtained to obtain the final detail-enhanced stylized images. The significance and generality of the algorithms and plugins proposed in this paper are verified through the comparison and analysis of relevant experimental data and SOTA methods.