Image style transfer has an important application in the field of art painting, and the existing art painting style transfer algorithms cannot well transfer the unique strokes and the contour features of the objects in the image of a specific-style work. In order to achieve better style transfer effect, a style transfer method for art painting images based on stroke and contour constraints is proposed. A pre-trained edge detector is used to extract the stroke features of the style image and the content image, and a loss function is designed to emphasize the stroke consistency between the two; erosion and blurring techniques are used to simulate the diffusion effect of pigments in oil paintings, and contour constraints are imposed on the object contours in the content image. The multi-head attention mechanism is introduced into the generative network to focus on key stylistic features, and the loss function adopts Smooth L1 instead of L1 to improve the training stability. The algorithm is mainly compared with CycleGAN in the experiments on the Van Gogh painting image dataset. The evaluation metric FID is reduced by 9%; SSIM is improved by 11.8%; PSNR is improved by 3.9%; and the human subjective evaluation is also significantly higher. The experimental results show that the algorithm has better style transfer effect than CycleGAN.