Study On Object Detection For High Resolution Remote Sensing Images Based On Support Vector Machines

The earth observation technique of high-resolution remote sensing (RS) is critical to national security and sustainable development. With the tendencies of "3-High" for RS images (high spatial resolution, high spectral resolution and high temporal resolution), the main developed countries pay more attention to the information processing technique for high-resolution image data. China proposed definitely in "tenth-five" 863 high-tech development plan: Developing advanced technique of acquiring high-resolution information in multi-dimension space, high-resolution aircraft-board optical and microwave data acquiring system for earth observation, ...Solving the key technique of high-precision and high-efficiency terrain classification, change detection for RS data processing and analysis. As a subtask of above, the object detection for high-resolution RS images has a good prospect in both military and commerce.However, the performances of current RS object detection methods and systems are not satisfying in both effect and efficiency. The main problems include: a) the data which the systems process has changed from high spectral resolution to high spatial resolution, which leads to an increasing input feature dimensionality and need a new method to deal with; b) with higher spatial resolution, the data size is exploring increasing and a new system is needed to solve the detection efficiency problem; c) to incorporate the new pattern recognition theories and methods into the field of object detection for new RS data.For the problems mentioned above, the main work of this paper is concluded as following:In section 1, a new classification method, support vector machines (SVMs), is applied. The object detection for high-resolution RS images is traditionally handled as 2-class classification problem based on empirical risk minimization principle, which generally leads to local optimal accuracy because of the limited training samples. The effect of object detection for high-resolution RS images is always not satisfying unless the size of samples is approaching infinity, which is hard to realize. The new theory, statistical learning theory (SLT), one the other hand, can systematically study the problem of machine learning with limited samples and present structural risk minimization principle, which leads to a good generalization ability for classifier. As the general applied method based on SLT, SVMs has an excellent performance in the learning problems with small size samples and high feature dimensionality.In section 2, a learning system based on appearance information of images is suggested to detect interested aircraft objects in RS images, which can build a detection model automatically and easily be extended to new objects. Based on the detection model, a SVMs classifier using pixel feature is constructed and achieves good experiment result. Four topics of interest are discussed here: a) pre-processing methods for training samples in pixel-based detection; b) comparing detection results of three common kernel models and relative parameters decision rules; c) a biasing training algorithm for unbalanced size of two classes samples. Experiment results with artificial data set areprovided here; d) an improved method to detect multi-oriented objects with mask pre-processing.In section 3, a wavelet method is presented for image representation. In common appearance-based detection, much pre-processing work is needed for every different type, quality images to achieve better result. The wavelet coefficient of objects is used here as a basic feature through Haar wavelet decomposition, which is easy to obtain multi-oriented and multi-scale edge information. Then, the selection, fusion and quantization of coefficients are discussed to avoid noise effect and improve performance. It shows that wavelet method can achieve an efficient and robust experiment result.In section 4, a bootstrapping strategy is proposed to solve sample selection problem. For RS object detection based on SVMs, the false alarm rate is usually hig...