The application of artificial intelligence in the field of art is becoming more and more widespread, especially in style transformation and art generation shows great potential. As an important form of artistic expression, traditional oil painting carries profound cultural connotation and aesthetic value. Aiming at the problems of unstable generation quality and insufficient texture feature extraction in traditional oil painting style migration, this study proposes an oil painting style migration method based on improved CycleGAN. The method adopts the relativistic discriminator and PatchGAN structure by introducing texture features as a priori knowledge input to the generator, and optimizes the loss function design. Experimental validation is carried out on a dataset containing 5000 images, and the Adam optimizer is used for 300 rounds of training. The experimental results show that the method outperforms existing methods in two evaluation indexes, structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR), where the SSIM value reaches 0.738 in Landscape→Oil style migration, which is an improvement of 0.127 compared to the GAN method. 800 oil paintings generated are evaluated, and the average score of the automatic evaluation is 4.236 points, and the average score of manual evaluation is 4.281 points. The texture characterization shows that different types of oil paintings present significant differences in indicators such as roughness and contrast. The study shows that the method effectively improves the quality and realism of oil painting style migration and provides a new technical support for digital art creation.
