UbiShop: Online Visually Similar Commodities Retrieval System on Mobile Phones
Date Issued
2010
Date
2010
Author(s)
Chi, Heng-Yu
Abstract
As the popularity of smart phones and the rapid evolution of mobile telecommunication, recently mobile phones provide users not only traditional communication function, but also integrated devices combined with advanced digital camera and developed wireless network. Since the capabilities of mobile phones are getting powerful, there are practical and interesting applications developed on mobile phones. Motivated by the fact that people usually cannot get the available information of commodities when seeing interesting ones accidentally, in this thesis, we develop an online visually similar commodity retrieval system named UbiShop that can provide related information of the desired commodity when users take a photo by the mobile camera as the input. For each commodity, this system also presents a list of visually similar images to recommend users similar commodities.
Moreover, we propose a real-time system framework that includes the components of mobile devices and the remote server. Regarding the search function on mobile phones, we design a real-time ranking method that is the faster descriptors matching algorithm based on pseudo hash-based indexing. Moreover, to achieve better accuracy performance, we adopt the RANSAC to further re-rank the top-k images generated from the proposed initial ranking method. Furthermore, to retrieve the similar commodities, we exploit the segmentation techniques to formulate the specific object-based mask which takes into account that feature comparison should be on different semantic regions respectively. A commodity is segmented into several regions that represent different semantic meaning respectively. Based on the mask, to recommend users similar commodities, we propose the Specific Object-based Regional Color Histogram matching (SORCH matching) and the Mask-SIFT matching to establish a recommended list of visually similar images. The proposed approaches are experimentally evaluated on a dataset of 8,000 watches images collected from the Amazon shopping website. The experimental results show that our system can provide useful information for users and can achieve satisfactory performance.
Subjects
multimedia mobile application
object retrieval
visually similar images
commodity recommendation
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-99-R97921026-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):ce6a1eced9a126c541a4c196f291dede