Visual product search in mobile business

• Main Author: Vijay, Dalmia Devanshu
• Other Authors: Yap, Kim Hui
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/61006
• Institution: Nanyang Technological University

Visual Search Technology allows users to retrieve information regarding visual objects. With the recent development of smartphones, this function can be performed on mobiles and is known as Mobile Visual Search (MVS). This report focuses on some of the image processing and pattern recognition techniques using the Mobile Visual Search. A mobile visual search application has already been developed on the Android platform using the client-server architecture. It uses the image processing techniques like Bag-of-Words model, Scale-invariant Feature Transform (SIFT) detector and descriptor, Inverted Index, Vocabulary Tree and Geometric Verification. One major part of this image recognition process is known as Keypoint Detection. The current application uses the SIFT (Difference of Gaussian) keypoint detector. This report seeks to evaluate some other keypoint detection techniques like Harris Affine and Hessian Affine. First, preliminary analysis is performed on the Harris Affine, Hessian Affine and the SIFT (DoG) detectors using the 48 image database provided by the Visual Geometry Group (VGG) at Oxford. These detectors are evaluated across five different image transformations which are viewpoint change, scale change, blur, light change and JPEG compression. Across all these transformations Hessian Affine is found to be the most optimal detector using the criteria which is explained in the chapter 4 of the report. Since the current Mobile Visual Search (MVS) application is in the C programming language, a C version of the Hessian Affine detector is found and is integrated in the current code and pipeline. The performance of this new MVS pipeline using the Hessian Affine detector is tested against the old pipeline which uses the SIFT (DoG) detector using the NTU Landmark Database. The percentage of images successfully recognized using the SIFT (DoG) detector is 84.36% whereas the percentage of images successfully recognized using the Hessian detector is only 81.60%. Contrary to the preliminary analysis, the SIFT (DoG) detector outperforms the Hessian Affine detector by 2.76%.