结合目标局部和全局特征的 CV遥感图像分割模型

摘 要：随着遥感卫星技术的发展，高分辨率遥感影像不断涌现。从含有较多信息、背景 复杂的遥感影像中自动提取目标成为一个亟待解决的难题。传统的图像分割方法主要依赖图像 光谱、纹理等底层特征，容易受到图像中遮挡和阴影等的干扰。为此，针对特定的目标类型， 提出结合目标局部和全局特征的 CV (Chan Vest)遥感图像目标分割模型，首先，采用深度学习 生成模型——卷积受限玻尔兹曼机建模表征目标全局形状特征，以及重建目标形状；其次，利 用 Canny 算子提取目标边缘信息，经过符号距离变换得到综合了局部边缘和全局形状信息的约 束项；最终，以 CV 模型为图像目标分割模型，增加新的约束项得到结合目标局部和全局特征 的 CV 遥感图像分割模型。在遥感小数据集 Levir-oil drum、Levir-ship 和 Levir-airplane 上的实 验结果表明：该模型不仅可以克服 CV 模型对噪声敏感的缺点，且在训练数据有限、目标尺寸 较小、遮挡及背景复杂的情况下依然能完整、精确地分割出目标。

CV image segmentation model combining with local and global features of the target

Abstract: With the development of the remote sensing satellite technology, high-resolution remote sensing images are on an increasing trend. The automatic target extraction from remote sensing images containing other information and complex background urgently needs to be realized. The traditional image segmentation method mainly depended on such underlying features as image spectrum and texture, and in image segmentation tasks, was likely to be impacted by the interference of shadow and occlusion in the image, complicating the segmentation and leading to unsatisfactory results. For this reason, according to the specific target type, a CV (Chan Vest) image segmentation model combined with local and global features of the target was proposed. Firstly, the deep learning generation model-CRBM (convolution restricted Boltzmann machine) was employed to represent the global shape features of the target and to reconstruct the shape of the target. Secondly, the edge information of the target was extracted by Canny operator, and a new shape constraint term integrating the local edge and global shape information was obtained by symbolic distance transformation. Finally, the CV model served as the image target segmentation model, and new constraints were added to gain the CV remote sensing image segmentation model integrating the local and global features of the target. The experimental results on the remote sensing dataset Levir-oil drum, Levir-ship and Levir-airplane show that the proposed model can not only overcome the noise sensitivity of the CV model, but also segment the target completely and accurately in the case of limited training data, small target size, occlusion and complex background.