Frequency and Space Domain Features for Image Classification Using Gaussian Mixture Models

This paper presents an effective combination of wavelet-based features and SIFT features, both of them have the frequency domain and space domain information characteristic. For the combined feature patches extracted from images we then adopt the PCA transformation to reduce the dimensionality of their feature vectors. And the reduced vectors are used to train Gaussian mixture models (GMMs) in which the mixture weights are adjusted iteratively. We experiment on Caltech datasets using this enhanced method, and the results comparing with several other methods show that the combination of salient feature vectors and GMM gives a much better improvement in image classification.