|关键词||高分辨率光学遥感图像 特征提取 旋转不变 特征编码 深度学习|
|英文摘要||Feature extraction for high-resolution optical remote sensing images is a hot topic in the remote sensing community. It is the basis of many computer vision applications in high-resolution remote sensing images, such as, object detection, scene classification, etc. With the development of computer vision and pattern recognition, feature extraction has made great progress in high-resolution optical remote sensing images. However, due to complicated and variable object appearances, dramatic illumination changes, objects with arbitrary orientations, partial occlusions and limited annotation data, extracting discriminative features from high-resolution remote sensing images is still a challenge problem.|
By combining prior knowledge with new methods in feature encoding and deep learning, this dissertation aims to develop feature extraction methods for object detection and scene classification in high-resolution optical remote sensing images. The main contributions can be summarized as follows.
1. A novel ordinal pyramid pooling method is proposed to extract rotation invariant features for objects in high-resolution remote sensing images. It first sorts pixels in the described region according to their radial gradient orientations (or radial gradient angles). Then a pyramid strategy is used to hierarchically partition these sorted pixels into even subgroups and local features of pixels are accumulated in each subgroup. This partition strategy can encode radial gradient orientation (or radial gradient angle) distribution information while maintain the rotation invariance, and improve the robustness to ordinal shift. The final representation is the concatenation of all pooled features from all subgroups. To maintain the rotation invariance, the proposed method uses Fourier HOG, a theoretically rotation invariant feature, as each pixel’s local feature and further improves its discriminative ability by Fisher Vector. Experiments on airplane detection and car detection in high-resolution remote sensing images demonstrate the effectiveness of the proposed method.
2. A novel visual feature description way named as rotation invariant matrix is proposed. It can incorporate angular spatial information in addition to radial spatial information, and maintain the rotation invariance. Based on the rotation invariant matrix, a rotation invariant feature extraction method with stronger discriminant ability can be proposed. Specifically, this method densely extracts rotation invariant matrix features from overlapping image subregions; then these extracted features are encoded into Fisher Vectors; finally, a pyramid pooling strategy that hierarchically accumulates Fisher vectors in ring subgroups is used to encode richer spatial information while maintaining the rotation invariance. These aborative designs help the proposed method achieve good experimental results on airplane detection and car detection in high-resolution remote sensing images.
3. A novel mixed-resolution encoded feature extraction method based on convolutional neutral networks is proposed. The low-level and middle-level intermediate convolutional features are respectively encoded by vector of locally aggregated descriptors (VLAD) and then reduced by principal component analysis to obtain hierarchical global features; meanwhile, the fully-connected features are average pooled and subsequently normalized to form new global features. The proposed encoded mixed-resolution representation is the concatenation of all the above global features. It can incorporate features from low-level, middle-level, and high-level simultaneously. Due to the usage of encoding strategies (VLAD and average pooling), the proposed method can deal with images of different sizes. Comparative experiments on two international public remote sensing scene classification datasets, UC-Merced and Brazilian Coffee Scenes, validate the effectiveness of the proposed method.
|王国利. 高分辨率光学遥感图像特征提取方法研究[D]. 北京. 中国科学院大学,2017.|