基于稀疏表示的旋转鲁棒遥感影像特征提取

Rotation Robust Remote Sensing Image Feature Extraction Based on Sparse Representation

针对高空间分辨率遥感影像的特征提取问题,提出一种基于稀疏表示的提取方法。通过学习,从大量的自然图像中获得过完备字典,对其中每个原子进行多个方向的旋转,从而扩展该字典。利用扩展的字典对遥感影像进行稀疏编码,并将稀疏编码非零元素个数限定为1,对非零元素的位置统计直方图进行池化处理后,通过归一化获得影像的特征。实验结果表明,与传统的特征提取方法相比,该方法可以有效提取遥感影像的特征,并且对遥感纹理影像的旋转具有较强的鲁棒性。

Aiming at the problem of high spatial resolution remote sensing image feature extraction, this paper proposes an extraction method based on sparse representation. It obtains overcomplete dictionary by learning on the mass natural images and then extends this dictionary by rotating every atom on multiple directions respectively. By using the extended dictionary, the sparse coding of the remote sensing image can be done, while the number of nonzero entries is limited to be 1. After the pooling process aiming to the location histogram of nonzero entries, it obtains the features of the image through normalization. Experimental result shows that compared with traditional feature extraction methods, this method can extract the features more efficiently and more robust about the rotation of the texture in remote sensing images.