Study On Key Techniques Of Content-based Retrieval Of Lunar Remote Sensing Images

Lunar remote sensing images are shot by the sensor equipments from the spacenear the moon surface, and they are the main data products and foundations for furthermoon studies. With the promotion of different missions in the current second lunarexploration boom, large amounts of lunar remote sensing images are accumulated. Howto retrieval these valuable data efficiently and thus make the best of them becomes oneof the new research hotspots. Content-based lunar remote sensing image retrieval,different from the traditional retrieval method based on metadata, no longer needs toprovide the complex and expertized different parameters as the retrieval input. Instead,several similar images in the lunar image database can be found as the retrieval resultsonly by analyzing and comparing the visual content of a query sample and images indatabase. It reduces the requirement for domain knowledge of users and makes theimage data available to more extensive users, which enables the fully use of lunarremote sensing images and brings out more scientific benefits. The research ofcontent-based lunar remote sensing image retrieval is significant both in theory and inpractice, which has caught the attention of researchers recently. Some efforts havealready been made on it globally, however, both the referring key techniques and thewhole retrieval method requires more improvements. Aiming at practical applications,this dissertation studies the problem of content-based lunar remote sensing imageretrieval by taking the idea of classical CBIR theory and techniques together withconsidering the domain specialty of lunar remote sensing images. The maincontributions of this dissertation include:(1) The problem of content features of lunar remote sensing imagesAccording to the specific visual appearance of lunar remote sensing images,considering the saliency of the image-contained craters, a salient region detectionalgorithm is proposed for lunar images. It merges and combines the highlight and darkSURF points to generate candidate ROIs, and through a SVM classifier the less salientor falsely detected ROIs are excluded from the final detection results. On the basis ofthe salient region detection, a kind of feature descriptors called LIFBS is proposed.LIFBS are calculated according to the comparative position, strength and scale relationsamong detected salient regions. The LIFBS generating algorithm is designed to supportmulti-threaded parallelism and is described in detail. Experimental results show that: theproposed detection algorithm is able to pick out most of the salient crater regions in alunar remote sensing image and works much better than the classical ITTi’s model. TheLIFBS describes the content of a lunar image well and is of good invariance andsimilarity. It works better than some classical global features such as grayscalehistogram, HU moments and Tamura texture descriptors. (2) The problem of high-dimensional indexes for feature vectorsFacing the NN search of global feature vectors, an exact high-dimensional indexsupporting concurrent queries called PCVAH is proposed. It uses hash-style structure toorganize the approximate vectors, filters the candidates by neighboring masks andsupports concurrent queries basing on multi-threaded parallelism. On the basis ofPCVAH, an approximate index called PQH is proposed. It partitions the query spaceand simplifies the filtering, which help to improve the NN search efficiency of globalfeature vectors. A PCPF index is proposed for matching the local feature vector (equalsto a2-NN problem) more efficiently. It uses a k-D tree to organize the query space,reduces the I/O by vector approximation and saves more time by a priority queue andthe restriction of the tracing nodes number. Experimental results show that: theproposed three indexes are of high query efficiency and outperform the classicalVA+-file, LSH and BBF index methods separately. For approximate methods, the queryprecision and recall of PQH is similar as that of exact search, while the matching resultby PCPF is better than that by BBF method.(3) The overall content-based lunar remote sensing image retrieval methodA compound feature model and a similarity measurement are proposed combiningthe contributions of the studies of feature descriptors and high-dimensional indexes forlunar remote sensing images. Basing on the feature model and the similaritymeasurement, the overall algorithm for similarity retrieval is proposed. After that, theoverall retrieval process is analyzed in detail and a Petri net tool is used to construct theprocess model. For parallel design in conditions of clusters and multi-cores, a parallelframework for content-based lunar remote sensing image retrieval is proposed byanalyzing the Petri net process model. Experimental results show that: the proposedmethod for retrieving lunar remote sensing images by the content is able to offervisually similar images in the database as retrieval results and the parallel frameworkhelps to speedup the overall process of retrieval.(4) The design and implemention of a prototype content-based lunar remotesensing image retrieval systemBasing on the lunar image database, a prototype system for content-based lunarremote sensing image retrieval is designed and implemented. On one hand, it verifiesthe effectiveness and feasibleness of the proposed algorithms. On the other hand, itdemonstrates the potential applications of the new retrieval method in the distributionand using of lunar exploratory data. The function, logical architecture and physicaldeployment of the prototype system are designed according to the analysis ofrequirements. Finally, the application of content-based lunar remote sensing imageretrieval is demonstrated via the prototype system.