Research On Oil Painting Style Generation Based On Deep Learning And Stroke Techniques

Image stylization is one of the research hotspots in the non-photorealistic rendering field.It can not only generate artistic works with distinctive styles but also preserve the visual key information of the target image,convey the key content that the artist wants to express through capturing the details of the work.Among the various styles,oil painting style has become one of the important visual forms of expression with its unique texture and visual appeal,and has great research value and wide application.In order to further improve the quality of stylized image oil painting,this thesis starts from the perspective of computer simulation artists realistically creating oil paintings,and proposes two improved brush-style oil painting stylization methods.The main contribution is summarized as follows:In response to the limitations,singularity,and time-consuming issues of traditional oil painting stylization methods that can simulate the real painting process,we propose a brush-style oil painting generation algorithm based on DCGAN-RBCA(DCGAN-Generator with Residual Block and Channel Attention)neural renderer by analyzing the artist’s real drawing process.According to the texture characteristics of oil paint accumulation,a parameterized brush structure with stripes was designed,and a differentiable DCGAN-RBCA neural renderer was constructed to generate strokes for the reference image.In addition,the stroke parameters were continuously optimized using gradient descent algorithm to minimize the difference between the generated image and the reference image.The results show that this algorithm can not only generate high-quality oil painting style images that fully perform the real painting process and details,but also has wide applicability,flexibility and adjustability.In response to the issue of the lack of brush texture in the oil painting image generated by Gatys’ style transfer method,we propose a convolutional neural network-based brush stroke oil painting style transfer algorithm.First,the DCGAN-RBCA neural renderer was used for brush-style reconstruction of content images,and then a pre-trained VGG19 network was used to define the style loss function of the style image,finally,the stroke parameters and style loss were simultaneously optimized to obtain a stroke-by-stroke generated oil painting style transfer image.The results show that the style transfer oil painting images generated by this algorithm have obvious brush textures and can generate high-quality oil painting effects in a brush style with style features.In order to facilitate user operation and test the practicality of the algorithm,this paper designed a brush-style system for generating stylized paintings.Users only need to upload input images,style images and select relevant parameter settings in the system.The system can then draw the corresponding stylized images,thereby transforming the theoretical content into a practical system that can be implemented.