Research On Order Based Feature Description For High-resolution Remote Sensing Image Recognition

High-resolution remote sensing images have been widely used in various fields, such as military interpreter, cartographic, urban planning, resource and environmental monitoring, agriculture and forestry monitoring, disaster monitoring, etc. With the development of earth observation technology, remote sensing images contain more and more information and will play an important role in much more fields. It’s a major issue that we need to deal with: How to automatically extract information and classify target of interest taking advantage of high-resolution remote sensing images. In this thesis, we research on the key technology of remote sensing image recognition using the theory of natural image recognition in computer vision. Specifically, we pay attention to the natural characteristic of high-resolution remote sensing images. It constructs feature descriptors based on order relation which provides quick and effective solutions for typical object recognition and general remote sensing image classification. The main contribution of this thesis can be summarized as follows: 1. We propose a light-weighted local feature descriptor called RBRIEF which uses only relatively few bits compared to other descriptors. The descriptor is constructed as a few binary bits from random binary comparisons. Unlike the original BRIEF, we use first derivative as a sample function to do binary comparisons which has been proven to be better compared against the function of intensity used in BRIEF. In addition, we discuss the selection of some parameters used in the construction of the descriptor and analyze the HASH principal contained in the random binary comparison. As a result, the performance of the descriptor outperforms SURF, BRIEF and ORB using standard benchmarks. 2. It does a parallel optimization for the entire construction process of RBRIEF and constructs an object recognition framework for high-resolution remote sensing images. We propose to compute integral image using a two-stage method based on the binary tree. Subsequently, we accelerate the computation of the gradient of the sample points in the process of RBRIEF construction using integral image and box filter. After that, we optimize the construction of RBRIEF based on GPU. At last, we discuss rapid object recognition based on RBRIEF. 3. We propose to use two intensity order based descriptors for classification of high-resolution remote sensing images. By analyzing the characteristics of remote sensing images, it points out that the imagery does not have an absolute reference frame, and can be seen as a fusion of natural image and texture image. For these characteristics, it uses two intensity order based approaches to extract low-level features of remote sensing images. The two descriptors are inherently rotation invariant and encode complementary information about the image. Experiments results on publicly available remote sensing imagery dataset show that MROGH performs better than SIFT and the combination of order based features which encodes complementary information improves the performance. 4. We construct a spatial co-occurrence matrix for soft quantization and perform a thorough evaluation of the effects of different classification frameworks(BoVW、SPMK、SPCK、SPCK+), encoding methods(VQ、LLC、KCB), kernel functions used in SVM(the linear kernel, histogram intersection kernel, additive chi-square kernel, exponential chi-square kernel) and the combinations of multiple kernels(mean, multi-kernel learning) by fixing the pipeline and its tuning based on remote sensing image dataset. In addition, The effectiveness of the spatial co-occurrence matrix has been proved by the experiment. 5. It gives a review of local image feature description. The thesis depicts the development history of local feature description in decades. Then based on the strategy of feature pooling, it classifies feature description methods into three types: histogram based method, feature comparison based method and machine learning based method. It gives a comprehensive overview of the various methods and compares them in terms of computational complexity, storage and descriptor performance. At last, the challenges and future development of local feature description has been discussed.