Learning Visual Dictionary Using Structured Sparse Principal Component Analysis
Popular visual category recognition techniques utilize the `Bag-of-Features' method, which represents an image as linear combination of the elements of a learned dictionary. The coefficients of the combination for an image constitute a feature vector which is typically high dimensional and sparse. The high dimensionality and the several uninformative elements in the dictionary together deteriorate classification performance. PCA is used to ameliorate the issues of dimensionality and very recently Sparse PCA (SPCA) has been utilized to compute a comparatively more informative dictionary. In our approach, we adapt Structure Sparse PCA (SSPCA), which leverages encoded higher-order information to build a better dictionary. We present a comprehensive empirical evaluation of our approach against SPCA on several popular datasets and different dictionary sizes.