计算机科学
生成语法
人工智能
图像压缩
自然语言处理
图像(数学)
生成模型
计算机视觉
图像处理
作者
Xianfeng Gu,Yuanyuan Xu,Kun Zhu
标识
DOI:10.1007/978-3-031-53308-2_6
摘要
Semantic image compression can greatly reduce the amount of transmitted data by representing and reconstructing images using semantic information. Considering the fact that objects in an image are not equally important at the semantic level, we propose a semantic importance-based deep image compression scheme, where a generative approach is used to produce a visually pleasing image from segmentation information. A base-layer image can be reconstructed using a conditional generative adversarial network (GAN) considering the importance of objects. To ensure that objects with the same semantic importance have similar perceptual fidelity, a generative compensation module has been designed, considering the varying generative capability of GAN. The base-layer image can be further refined using residuals, prioritizing regions with high semantic importance. Experimental results show that the reconstructed images of the proposed scheme are more visually pleasing compared with relevant schemes, and objects with a high semantic importance achieve both good pixel and semantic-perceptual fidelity.
科研通智能强力驱动
Strongly Powered by AbleSci AI