A New Unsupervised Feature Learning Method for Object Recognition using Prior-Knowledge Data

In this paper we first propose a Self-Organized Incremental Gaussian mixture based clustering mechanism (we named it as SOIGmm). The recognition system runs this clustering mechanism on some local features, which have been extracted, from some prior unlabeled data to create a strong visual codebook. In addition, a wavelet transform is applied for extracting color-based features. The second objective is considering the spatial importance of the different parts of each image into the representation. The saliency map is extracted and used which has a high role in the representation of each image. Furthermore, in the classifier, we used pyramid matching to find similarity between images and used averaging to combine the kernels on the classifier. The experimental results show that the proposed method could achieve excellent performance on object recognition both for one shot and multi-shot training data sets.