论文部分内容阅读
遥感图像包含的信息丰富,纹理复杂,而遥感图像语义标注又为后续的目标识别、检测、场景分析及高层语义的提取提供了重要信息和线索,这使其成为遥感图像理解领域中一个关键且极具挑战性的任务。首先针对遥感图像语义标注问题,提出采用条件随机场(CRF)框架对遥感图像的底层特征和上下文信息建模的方法,将Texton纹理特征与CRF中的自相关势能结合来捕捉遥感图像中的纹理信息及其上下文分布,采用组合Boosting算法进行Texton纹理特征选择和参数学习;然后将Lab空间中的颜色信息与CRF中的互相关势能结合来描述颜色上下文;最后用Graph Cut算法对CRF进行推导求解,得到图像自动语义标注结果。同时,建立了可见光遥感图像数据库Google-4,并对全部图像进行了人工标注。Google-4上的实验结果表明:采用CRF框架与Texton纹理特征和颜色特征相结合对遥感图像建模的方法与基于支持向量机(SVM)的方法相比较,能够取得更准确的语义标注结果。
Remote sensing images contain abundant information and complex textures. Semantic annotation of remote sensing images also provides important information and clues for the subsequent target recognition, detection, scene analysis and high-level semantic extraction, making it a key in the field of remote sensing image understanding Very challenging task. Firstly, aiming at the problem of semantic annotation of remote sensing images, this paper proposes a method of using the conditional random field (CRF) framework to model the underlying features and contextual information of remote sensing images. Textur texture features and the auto-correlation potential in CRF are combined to capture the texture of remote sensing images Information and its context distribution, using the combination of Boosting algorithm Texton texture feature selection and parameter learning; then the color space in Lab space and CRF cross-correlation potential to describe the color context; Finally, Graph Cut algorithm for derivation CRF , Get the image automatically semantic annotation results. At the same time, Google-4, a visible light remote sensing image database, was established and all the images were manually annotated. The experimental results on Google-4 show that using the CRF framework combined with Texton texture features and color features to model remote sensing images can achieve more accurate semantic annotation results than SVM-based approaches.