Studying The Martial Object Detection For High-resolution Remote Sensing Images

With the development of the technology, modern remote sensing can help us observe and measure the ground objects with a dynamic, fleet, accurate and multi-means way. And the quality of remote sensed images tends to be high spatial resolution, high spectral resolution and high temporal resolution. So the remote sensing has been wildly applied in many fields such as agriculture, environmental monitoring, economic development, national defense and so on. The technology of object automatic detection in remote sensing image with high resolution is a very important research field in recent years, which means the useful target in the remote sensing images can be located and recognized by some kind of technological means. The object detection for high-resolution RS images has a good prospect in both military and commerce.Machine leaning which is based on data is the most important part of the object detecting system which is based on computer vision and pattern recognition Support Vector Machine (SVM) is a kind of novel machine learning method developed by Vapnik and his research group. Due to its perfect learning performance, this technique has become hot in the field of machine learning .It also has successful applications in many fields. Compared with theoretical research, the application research in remote sensing image is quite slow.1. A classification method, support vector machines(SVMs) is applied. Considering of large training set, there existing some redundancy, and the classifying efficiency and correct rate being only related with support vectors(SVs), we propose a new SVM iterative algorithm based on K-mean clustering algorithm. K-mean clustering algorithm is used to compress the training set and the new original training set is gained. Then the margined samples and error-classifying samples are joined in the new original training set to renew it. This process is iterative till the false samples'number is not changed. So the speed of learning is accelerated while keeping the same train precision, and the speed of classification is also improved.2. Rapid detection algorithms are put forward to efficiently detect object in broad area RS image. For large image with complex patterns, the common classifier based on SVMs results in large size of support vectors, which will decrease the detection efficiency. In this section, one fast search strategy is discussed and its corresponding algorithm is presented either. A bottom-up algorithm based on hierarchy of SVMs is used to improve the speed of detection, which is constructed by several SVMs classifiers whose complexity is gradually increase. The algorithm can perform efficiently by rapidly removing the obvious background and reducing the number of regions that are provided to detect by more complex classifiers.3. Considering the lack of target' geography information, combined with the texture analysis of remote sensing image, we propose a fast getting ROI's method based on SVM classification networks. By dividing the class of non-ROI into three subclasses, we construct three SVMs between three subclasses and the class of ROI. Then, the three SVMs construct a classification network. We can use this network to getting ROI. Aimed at ROI's automatic target multi-classification, a binary decision tree based on SVM is proposed. Compared with one SVM method, this network classification methods has faster training speed and higher correct rate. We get good result when we applied it to the airplane detection in lack of target' prior knowledge.