Improving image classification by orthogonality of sparse codes
Céline Rabouy, Sébastien Paris, Hervé Glotin · 2015
Sparse Coding (SC) is an approach widely used in image classification. It allows to reconstruct the signal with few elements and follows the specific scheme of Bag-of-Words (BoW). However, we can observe a decorrelation between input patches and reconstructed patches. To answer that, Graph regularized Sparse Coding (GSC) exists. As GSC works on the training set, we propose a new modeling, Joint Sparse Coding (JSC), for the testing set. JSC can be seen as a tradeoff between SC and GSC. To go furthermore, we explore the simple fusion of models. To explain the observations of the fusion results, we will be led to study the orthogonality properties by the cosine computation. These applied on UIUCsports, 17Flowers and scenes15 lead us to put forward the various qualities of the studied bases and sparse representation. We demonstrate a significant improvement of the State-of-the-Art for the UIUCsports database.