Image style migration is a research hotspot in the field of digital media art and artificial intelligence computer vision. The automatic style migration method based on transfer learning (TLST) proposed in this paper undergoes style extraction, style learning and style migration, obtains the style feature matrix, calculates the correlation between the features, and completes the remapping of the content of the original style image. The efficiency of the TLST algorithm is verified by comparing with other models and the application study on real devices. Specific results show that the Precision, Recall and F1 values of TLST are above 0.8, thus, TLST outperforms several other neural network models. The method in this paper reaches a maximum accuracy of about 100% after several trainings, which is better in artist painting recognition. Deploying TLST on MSP432P401R and ARM CortexM7 platforms reduces the average inference time from 156.90 seconds to 146.80 seconds and from 63.021 seconds to 59.324 seconds, respectively. This shows that TLST is able to reduce redundant computation and decrease inference latency.
