BoVW (Bag of Visual Words) Model has attracted much attention for many computer vision applications in which an image is represented by a histogram of visual words. Two of its critical steps are to construct a visual dictionary and to quantize each local feature to its nearest visual word in the dictionary. In this paper, we present the framework of a generalized BoVW (GBoVW) Model in which feature quantization can be replaced by sparse coding based feature encoding. We also propose to use spectral clustering to construct a visual dictionary to overcome the shortcomings of K-Means based clustering algorithms. Image retrieval experiments on ZuBud database indicate that GBoVW Model improves BoVW Model and the visual dictionary generated by spectral clustering achieves better performance than that by K-Means based clustering methods.
Bag-of-Visual Words and Error-Correcting Output Codes for Multilabel Classification of Remote Sensing Images