基于稀疏表示分类的人工地物目标检测

Man-made Object Detection with Sparse Representation-based Classifier

针对遥感图像中人工地物目标复杂性和多样性的问题,提出了一种基于稀疏表示分类的人工地物目标检测方法。根据遥感图像的特点,首先,在冗余Contourlet变换域中对遥感图像进行了预处理,降低了噪声的干扰;其次,研究了高效的特征提取方法,通过在冗余Contourlet多级分解中引入最优基函数选择策略,计算出遥感图像的旋转不变纹理特征和基于低频的强度特征,并与遥感图像中的分形误差特征进行组合,得出复合特征向量;最后,利用稀疏表示分类方法对提取出的组合特征进行处理,完成了人工地物目标分类,并且利用数学形态学操作对分类的结果进行了优化。实验结果表明,该方法对人工地物目标检测具有较好的鲁棒性和准确性。

Man-made object detection is challenging due to the complexity and variability of aerial imaging.In order to solve this problem,a novel man-made object detection method based on sparse representation classifier is proposed.Considering the characteristics of aerial images,the acquired images were firstly pre-processed to reduce noise disturbance based on the non-subsampled Contourlet transform.Then an efficient feature extraction method was investigated.The rotation-invariant texture feature and intensity feature were calculated through the best basis selection strategy,which was applied in the non-subsampled Contourlet decomposition process.Then,these features were further combined with the fractal error features.At last,the sparse representation-based classifier was employed to deal with the combined features to acquire the classification results,which need to be further refined by the morphological operations.The experimental results show that the proposed method is robust and accurate.