A Semantic Enhanced Linear Coding for Image Classification

Considering the ambiguity problem in the traditional feature coding model, a feature context-aware semantic enhanced linear coding method is proposed. At first, the context information is represented by the concurrence matrix learnt from local area of the features. Then, the context information and a context matching weight are introduced into the coding model to form a new semantic enhanced coding model. Owning to the functions of context information and the context matching weight, this model not only enriches the semantic meaning of coding, but also efficiently avoids the affects of noise. Experiments on the baselines (Scene15, Caltech101, and Caltech256) demonstrate the effectiveness of the proposed method.