Shape Classification Using Local and Global Features

In this paper, we address the shape classification problem by proposing a new integrating approach for shape classification that gains both local and global image representation using Histogram of Oriented Gradient (HOG). In both local and global feature extraction steps, we use PCA to make this method invariant to shapes rotation. Moreover, by using a learning algorithm based on Adaboost we improve the global feature extraction by selecting a small number of more discriminative visual features through a large raw visual features set to increase the classification accuracy. Our local method is adopted from the popular bag of key points approach for shape classification. To integrate the classification results generated based on both local and global features, we use a combining classifier to perform the final classification for a new unknown image query. The experiment results show that this new method achieves the state-of-art accuracy for shape classification on the animal dataset in [8].