Predictive high-level feature representation based on dictionary learning

A much improved computational performance of visual recognition tasks can be achieved by representing raw input data (low-level) with high-level feature representation. In order to generate the high-level representation, a sparse coding is widely used. However, a major problem in traditional sparse coding is computational performance due to an ℓ0/ℓ1 optimization. Often, this process takes significant amount of time to find the corresponding coding coefficients. This paper proposed a new method to create a discriminative sparse coding that is more efficient to compute the coding coefficients with minimum computational effort. More specifically, a linear model of sparse coding prediction was introduced to estimate the coding coefficients. This is accomplished by computing the matrix-vector product. We named this proposed method as predictive sparse coding K-SVD algorithm (PSC-KSVD). The experimental results demonstrated that PSC-KSVD achieved promising classification results on well-known benchmark image databases. Furthermore, it outperformed the currently approaches in terms of computational time. Consequently, PSC-KDVD can be considered as a suitable method to apply in real-time classification problems especially with large databases.