Large-scale image collection cleansing, summarization and exploration
1 online resource (206 pages) : PDF
University of North Carolina at Charlotte
A perennially interesting topic in the research field of large scale image collection organization is how to effectively and efficiently conduct the tasks of image cleansing, summarization and exploration. The primary objective of such an image organization system is to enhance user exploration experience with redundancy removal and summarization operations on large-scale image collection. An ideal system is to discover and utilize the visual correlation among the images, to reduce the redundancy in large-scale image collection, to organize and visualize the structure of large-scale image collection, and to facilitate exploration and knowledge discovery.In this dissertation, a novel system is developed for exploiting and navigating large-scale image collection. Our system consists of the following key components: (a) junk image filtering by incorporating bilingual search results; (b) near duplicate image detection by using a coarse-to-fine framework; (c) concept network generation and visualization; (d) image collection summarization via dictionary learning for sparse representation; and (e) a multimedia practice of graffiti image retrieval and exploration.For junk image filtering, bilingual image search results, which are adopted for the same keyword-based query, are integrated to automatically identify the clusters for the junk images and the clusters for the relevant images. Within relevant image clusters, the results are further refined by removing the duplications under a coarse-to-fine structure. The duplicate pairs are detected with both global feature (partition based color histogram) and local feature (CPAM and SIFT Bag-of-Word model). The duplications are detected and removed from the data collection to facilitate further exploration and visual correlation analysis. After junk image filtering and duplication removal, the visual concepts are further organized and visualized by the proposed concept network. An automatic algorithm is developed to generate such visual concept network which characterizes the visual correlation between image concept pairs. Multiple kernels are combined and a kernel canonical correlation analysis algorithm is used to characterize the diverse visual similarity contexts between the image concepts. The FishEye visualization technique is implemented to facilitate the navigation of image concepts through our image concept network. To better assist the exploration of large scale data collection, we design an efficient summarization algorithm to extract representative examplars. For this collection summarization task, a sparse dictionary (a small set of the most representative images) is learned to represent all the images in the given set, e.g., such sparse dictionary is treated as the summary for the given image set. The simulated annealing algorithm is adopted to learn such sparse dictionary (image summary) by minimizing an explicit optimization function.In order to handle large scale image collection, we have evaluated both the accuracy performance of the proposed algorithms and their computation efficiency. For each of the above tasks, we have conducted experiments on multiple public available image collections, such as ImageNet, NUS-WIDE, LabelMe, etc. We have observed very promising results compared to existing frameworks. The computation performance is also satisfiable for large-scale image collection applications. The original intention to design such a large-scale image collection exploration and organization system is to better service the tasks of information retrieval and knowledge discovery. For this purpose, we utilize the proposed system to a graffiti retrieval and exploration application and receive positive feedback.
Ribarsky, WilliamRas, ZbigniewXiao, JingZheng, Nigel
Thesis (Ph.D.)--University of North Carolina at Charlotte, 2012.
This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s). For additional information, see http://rightsstatements.org/page/InC/1.0/.
Copyright is held by the author unless otherwise indicated.