随着遥感技术的发展，遥感影像信息变得越来越丰富。面对日益增长的遥感数据规模，人们当下研究的热点不约而同地聚焦到了如何能够准确而快速地完成遥感影像的分析任务。 近些年来，以视觉注意机制为代表的显著区域检测技术被引入到了遥感影像的分析领域，在大规模数据亟需高效分析的背景下成为了一项提高数据处理实时性和准确性的重要技术方法。它可以给观察者提供遥感影像的显著区域 (Salient Region)，或称感兴趣区域 (Region of Interest，ROI) 的信息，继而在该区域中寻找具体目标。通过这样的方式可以帮助制定更为合理的计算资源分配方案，从而大幅地提升己有遥感影像处理系统的运行效率。目前，学者们提出的感兴趣区域检测模型主要是基于认知、信息论、频域、图论等，且都是对单幅图像进行处理。然而，考虑到显著性分析在目标联合分割和联合识别，以及在视频领域等方面的应用，联合显著区域检测正在悄然兴起。 通常，卫星为了准确地获取地面遥感信息，会在相同或者相近位置连续拍摄多张图像，由此产生了大量具有相似特征的高空间分辨率遥感影像。虽然采用针对单幅图像的显著性提取算法，也能得到它们的显著图，但是这种方式忽略了各图像间的内在相似性，因而提取效率和精度都不高。 本文针对以上问题，提出一种基于联合显著性分析的高空间分辨率遥感影像感兴趣区域提取模型。我们将联合显著性引入这些待处理的遥感影像组，正是利用了各图像同属一个图像源，且各显著区域相似度高的特点，达到同时提取一组图像的显著区域的目的。该模型首先采用二分 K-means 在 RGB 和 CIELab 两种颜色空间上对所有图像进行聚类，以便后续过程能在簇的层级计算显著值，从而达到降低计算复杂度的目的。然后，我们提出了 LabH 颜色直方图，并结合面积-周长比来综合有效地进行全局对比度计算。接着，通过融合两种不同颜色空间的显著图，以进一步去除复杂背景中小路和阴影，得到突出城镇居民区域的联合显著图。最后，在感兴趣掩膜中加入 gPb 边缘检测来限制显著区域内部的形态学孔洞填充，并保持边缘清晰，以提高本模型的性能。 为了评价提出模型的性能，我们选取了三组不同分辨率的遥感影像（10 幅 512 × 512，14 幅 1024 × 1024 和 10 幅 2048 × 2048），从定性和定量两个角度与八种当今最先进的显著区域检测方法进行了实验比较。无论是主观的显著图和显著区域检测结果比较，还是客观的ROC（受试者工作特征曲线）和PRF（查准率、查全率和 F 值）的评价指标比较，我们提出的模型均优于另外八种显著区域检测模型。实验证明，本文对今后的高空间分辨率遥感影像的感兴趣区域提取具有重要的实际意义。

With the development of remote sensing technology, current high-resolution images contain more information and have been widely used in various fields. To process the rapidly growing amount of data in remote sensing images, a more efficient information processing technology is urgently needed. Salient region detection technology, which is represented by the visual attention mechanism, has been introduced into the remote sensing image analysis field, and it has become an important technical approach for improving the computation efficiency and analysis accuracy in mass-data image processing. After providing a salient region, or region of interest (ROI), the viewer can search for specific objects in the region. By this way, it can help to develop a more rational allocation of computing resources and significantly enhance the operating efficiency of an image processing system. Several computational models based on visual perception, information theory, frequency domain, graph theory and so on, have been developed and primarily processed on a single image. However, giving the wide application prospect of joint segmentation, joint identification and applications in video area, co-saliency is well worth studying. Typically, satellite will take multiple images continuously at the same position or somewhere nearby to obtain adequate information. Thus, it produces a lot of high-resolution images with similar characteristics. Although the use of the algorithms designed for single image can extract saliency maps as well, these approaches ignore the inherent similarities between images. Therefore, the extraction efficiency and precision are both decimated. To tackle these problems, we propose a salient region detection model based on co-saliency for high-resolution remote sensing images. Since all images are from a single source and share high similarity, by applying the co-saliency strategy, we can simultaneously get a set of salient regions for all the images. The model firstly adopts the bisecting K-means to cluster all the images in the RGB and CIELab color space, respectly, which makes it possible to calculate saliency on cluster-level and thus greatly reducing the computational complexity. Then, we join LabH color histogram with area-perimeter ratio to conduct the global contrast between these clusters in an efficient way. Furthermore, we manage to remove the interfering paths and shadow in the complex background, while only keep the residential area in the downtown as salient, through the integration of saliency map born in different color spaces. Finally, gPb edge detection is added to limit the morphological area filling operation within the interior of each salient area, and at the same time, the sharp edges are well preserved. Hence, the performance of the model is improved overall. We compare our algorithm to eight state-of-the-art salient region detection methods qualitatively and quantitatively on three groups of remote sensing images of different resolution (ten for 512 × 512, fourteen for 1024 × 1024 and ten for 2048 × 2048). It turns out that our method not only outperforms the other algorithms visually, but also achieves better ROC and PRF in the experiments, which further proves that our model has certain practical value for the saliency detection of remote sensing images.