Online sparse learning utilizing multi-feature combination for image classification
Lihe Zhang, Kunyu Zhang, Xiaoli Dong · 2011
Bag-of-features has become very popular in Image classification. Offline codebook learning has to limit the number of training sample concerned with memory, and it influences classification accuracy to some extent. We propose an online sparse learning algorithm, which utilizes the reconstruction error to update the current codebook. It can capture salient properties of images in real-time. Most of image representation approaches in Gabor domain merely utilize magnitude information, and some important phase information is missing. Taking both magnitude and phase response into account, a Local Gabor Magnitude Weighted Phase (LGMWP) descriptor is proposed in this paper. The technique works by dividing the image into local patches, extracting SIFT and LGMWP features to online learn the codebook respectively, implementing spatial pyramid matching (SPM) and binary SVM classifier. The experiment results demonstrate our approach outperforms offline learning with a single type of descriptors.