UbiShop: A Ubiquitous Image Search System for Online Shopping Commodity
Date Issued
2010
Date
2010
Author(s)
Kung, Chien-Ray
Abstract
As advancement of multimedia analysis and machine learning proceeds with each passing day, it becomes more possible to recognize captured images, and retrieve similar images and related information by comparing extracted features of query image with large numbers of images in datasets. In addition, thanks to popularity of smart phone and evolution of mobile telecommunication grow rapidly, people are gradually able to use mobile device to get desired information ubiquitously through transmission from remote servers. Combining these trends, we propose an online shopping commodity image search system, UbiShop, which makes users snap interesting commodities through smart phones and receives information about the commodities from the web.
In this thesis, to solve problem of previous local descriptors matching with high cost time, we propose Hierarchical pseudo Hash-based Descriptors Matching (H-pHDM)
algorithm that greatly speeds up traditional descriptors matching with little accuracy loss. The experiments conducted on different testing datasets achieve great performance. Furthermore, to exploit to UbiShop system and concern about both retrieval time and accuracy, we design a two-level ranking mechanism that keeps top-K images which are most likely the matched images by H-pHDM algorithm, and re-ranks these images by conventional local descriptors matching criteria.
Besides, we even construct similar images lists for watches datasets that recommend users other watches with similar visual style as target watch is retrieved. Hence, we propose Specific Object-based Regional Color Histogram (SORCH) feature matching approach that compares color histogram on different regions of objects respectively and ranks by similarities with other images to construct the lists.
In addition, considering the limitation of computation power and capacity of current mobile devices, we design a client-server architecture system to put complex computation to server side, yet keep simple user interaction on mobile device. The notion lets large numbers of images and information be stored in the server and makes the system be extensible with more commodity categories in the future.
Subjects
Image Retrieval
Multimedia Indexing
Machine Learning
Mobile Application
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-99-R97921038-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):44333541c49afebfbd5d8c7c12fde3a6