A Hybrid CNN-GrabCut and GAN Architecture for Semantic Segmentation and Artistic Style Transfer in Image Generation

Abstract

Driven by digital transformation and the growing demand for artistic innovation, the development of artistic image generation systems has gained wide attention. However, existing systems face several limitations, including low accuracy in semantic segmentation, unnatural fusion in style transfer, and weak originality in generated images. Therefore, an artistic image generation system model based on graph cutting method convolutional neural network generative adversarial network is proposed, which uses an improved convolutional neural network (CNN) combined with GrabCut algorithm for semantic segmentation. By introducing convolutional block attention module to optimize CNN, it can better capture global information and complex features. In terms of style transfer, the system uses a semantic guided GAN architecture to achieve precise adaptation and natural transition of style features. The study achieved a pixel accuracy of 98% on the PASCAL VOC dataset. With the increase of training data, its highest intersection to union ratio reached 92%, significantly higher than the comparison algorithm. In real-time testing, the system only occupies 710 MB of memory and has a response time of less than 62 ms, outperforming other models in performance. In addition, in the image quality test conducted on the ArtBench-10 dataset, the system achieved a peak signal-to-noise ratio of up to 54 dB and a structural similarity index of 92%. These results indicate that the proposed model delivers high accuracy and strong diversity in painting art design. It effectively solves current problems in segmentation precision and style fusion, offering new ideas for artistic creation and supporting the development of intelligent painting systems.