Region-based Multi-resolution Remote Sensing Image Semantic Retrieval

In recent years, with large amount of satellites blastoff and sensor technology rapidly development, the quantity of remote sensing images in the database exponentially increases. Although, the development of network technology, database technology and software technology greatly improves the ability of image organizing, managing and processing, it is still difficult to quickly and effectively retrieve interested images due to huge and unstructured remote sensing image database. Content-based image retrieval system emerged and developed to solve this problem. However, there exist two main problems. Firstly, the difference between low-level image feature and high-level semantic feature makes precision of image retrieval decrease. This difference is named "semantic gap". How to mine the hidden semantic features of images becomes the hotspot in image retrieval research fields. Secondly, remote sensing image is different from ordinary picture. Remote sensing images which come from the same area but with different resolution also contain different spectrum, texture and sharp features, which makes it difficult to retrieve between different resolution images.In this paper, researches are focused on how to narrow semantic gap of remote sensing images and how to realize multi-resolution remote sensing image retrieval. The main contributions of this thesis are as following:Region-based image retrieval is aimed to narrow the semantic gap. So, in this research image segmentation will be finished at first step. Complicated spectrum and texture feature and random factors which are brought during image screening will make image segmentation more difficult. JSEG algorithm is first imported to segment remote sensing images in this research, and it is improved to be applicable to segment multi-spectral remote sensing images. Then, original spectral feature is replaced by NDVI, NDBI and mean texture for segmentation. Experiments demonstrate that JSEG algorithm present a good segmentation result not only for images which contain complicated texture feature but also for which contain complicated spectral feature, and it avoids oversegmentation.Semantic feature extraction is another method of narrowing semantic gap. Semantic features are extracted by using semantic level model. The basic three levels semantic model contains feature level, object level and scene level. But, it’s difficult to extract object basing on remote sensing image features. So, a new three-level remote sensing image content representation model is built in this research. Based on this model, region segmentation and region feature extraction, semantic feature is extracted by Expectation Maximization method. Semantic feature experiments demonstrate that semantic features can not only replace low-level image feature and also well represent image content. Image retrieval experiments demonstrate that semantic-based image retrieval can get a better retrieval precision.Relevant feedback technology also can narrow semantic gap, but increases the complexity of image retrieval. In this research, positive example and negative example are used to replace feedback. This way not only decreases the complexity of image retrieval but also increases retrieval precision.Remote sensing images which come from same area but different resolution contain different texture features and spectral features. This makes multi-resolution remote sensing image retrieval more difficult. In this research, super-resolution reconstruction technology is used to solve this problem. Low resolution image is reconstructed to high resolution image, and then high resolution image is used to retrieve, and this solves multi-resolution image retrieval problem to a certain extent. Experiments demonstrate that this method can get better results.