Using Sparse Representation Classifier (SRC) to Calculate Dynamic Coefficients for Multitask Joint Spatial Pyramid Matching

Using multiple feature descriptors simultaneously increases the accuracy of object recognition. Usage of dynamic coefficients, in other words, non-identical and proportional to the importance of each descriptor for each class, is an appropriate and efficient way to combine different feature descriptors. In this paper, a new and efficient structure is proposed that calculates these coefficients using the sparse representation classifier. For each feature descriptor, we propose an important criterion based on the reconstruction error of the images via the sparse representation. The assigned importance of each descriptor for each class is different and calculated based on the reconstruction errors of the images when only their classmate images contribute in the reconstruction process. In addition, an innovative method is proposed which can be used to help classes that are not well described by any descriptor as well. In this method, using the residual criteria, these classes are identified and using a defined notion of similarity among classes, the accuracy of these classes with the support of the similar ones is enhanced. The experimental results of the proposed work on Caltech-101 and Caltech-256 databases show the success of approaches compared with state-of-the-art ones on the same databases.