Object Recognition Based on the Region of Interest and Optimal Bag of Words Model.

Weisheng Li,Peng Dong,Bin Xiao,Lifang Zhou
DOI: https://doi.org/10.1016/j.neucom.2015.01.083
IF: 6
2016-01-01
Neurocomputing
Abstract:Bag of Words (BoW) model has been widely used in conventional object recognition tasks. Different from the existing methods, this paper proposed a method for object recognition based on Region of Interest (ROI) and Optimal Bag of Words model. It includes the following steps: (1) ROI extraction in combination with the Shi–Tomasi corner and Itti saliency map; (2) The SIFT feature descriptors are detected and described on images of interest; (3) A visual codebook is generated through utilizing the Gaussian mixture models, which relies on the clustering results of k-means++; (4) The similarities between each visual word and corresponding local feature are computed by posterior pseudo probabilities discriminative to construct a visual word soft histogram for image representation; (5) The Support vector machine (SVM) is used to perform image classification and recognition. The experiments are performed on the MSRC 21-class database. The results show that the proposed method can be more accurately recognize images.
What problem does this paper attempt to address?