부산대학교 도서관

Semantic information in images can be extracted using semantic segmentation, an essential task in the field of computer vision. Semantic segmentation, however, faces several challenges across a wide variety of domains, including textiles, cartoons, and real images. The semantic segmentation models need better contextual information comprehension and target segmentation abilities due to the complexity and abstraction of patterns in the textile domain, the variety of cartoon drawing styles, and the complexity and variety of image backgrounds in real scenes. In this work, we first define the semantic labels of patterns and construct multi-domain datasets such as textiles, cartoon images, and daily scenes; then, based on the DeepLabV3+ model, we propose the optimized model MDE-DeepLabV3+, which introduces the channel spatial attention mechanism to obtain more feature information and improve network performance ability. The experimental results show that the mean Intersection over Union of this paper's method on the multi-domain dataset is 79.00%, and when compared to the traditional method, this paper's model effectively improves the accuracy and richness of feature extraction while reducing feature detail information loss.