Supervised Incremental Feature Coding for SAR Image Classification

Abstract

Recent years have witnessed a fruitful development of image representation. The most prevalent one is the Bag-of-Words (BoW) method, which gives state-of-the-art performance in many applications. It has four steps: local feature extraction, dictionary learning, feature coding, and feature pooling. In this paper, we focus on feature coding. On the basis of an analysis of currently popular feature coding methods, we propose a supervised incremental coding method. The most different characteristic of this method is that coding of a new image relies on the coding of the previous image from the same class. Therefore, we need to know the label of one image before coding. This point can be argued as a drawback of this method. However, we demonstrate that it could give much better feature for image classification. This finding gives some hints about further development of feature coding method. We believe that the entire class should be considered when coding the local features.