Sparse Representation Methods for Image Prediction

Abstract

An alternative class of image prediction methods based on sparse representation is discussed in this chapter. Typical sparse representation techniques aim to approximate the unknown block as a linear combination of few patches extracted from an adaptive dictionary, that is formed by texture patches derived from the previously encoded area. Due to their features, sparse representation methods tend to perform better in the presence of complex textures with repeated patterns. The sparse characteristic imposes a reduced number of linearly combined patches, providing a representative model that performs more efficiently for prediction purposes.

In this chapter, dimensionality reduction methods are investigated in the context of the most recent HEVC standard, in particular for Light Field image coding. Experiments demonstrate the advantage of the LLE-based prediction to exploit the high redundant structure of Light Field images.

An improved sparse linear prediction method that can be seen as a generalisation of existing linear prediction and sparse representation prediction methods is also presented. Unlike most LSP-based algorithms proposed in literature, which use a fixed filter context limited to a small set of closer neighbouring pixels, sparse-LSP involves a linear prediction method able to use a sparse filter context defined in larger causal area. Sparsity restrictions are imposed to the linear prediction method by limiting the number of non-null coefficients in the filter context, using the k-nearest neighbours method. Experimental tests for natural image coding demonstrate the advantage of the developed method relative to other existing prediction solutions.