An Object Recognition Algorithm Based on Kernel Function Matching in Spatial Pyramid

摘要

Aiming at the shortcomings of traditional object recognition algorithms based on Bag of Words (BOW) model, we present an object recognition algorithm based on kernel function matching in spatial pyramid. Firstly, we extract Dense Scale-invariant Feature Transform (DSIFT) features of objects and cluster the DSIFT features of training samples by K-means++ algorithm to get visual words. Secondly, we map the DSIFT features and visual words to a high dimensional space in spatial pyramid and use kernel function matching which constructs the histogram of visual words for each object more accurately. Finally, we use Support Vector Machine (SVM) classifier which is extended to multi-class decisions by combining one-versus-all votes to complete the task of object recognition. The experiments are carried out in two classical computer vision databases: Caltech 101 and Caltech 256. The results show that our algorithm performs better than traditional methods. What is more, the recognition accuracy is good even in the case of small samples.