Histogram of Structure Tensors: Application to Pattern Clustering

Pattern clustering is an important data analysis process useful in a wide spectrum of computer vision applications. In addition to choosing the appropriate clustering methods, particular attention should be paid to the choice of the features describing patterns in order to improve the clustering performance. This paper presents a novel feature descriptor, referred as Histogram of Structure Tensors (HoST), allowing to capture the local information of an image. The basic idea is that a local pattern could be described by the distribution of the structure tensors orientations and shapes. The proposed HoST descriptor has two major advantages. On the first hand, it captures the dominant orientations in a local spatial region taking into account of the local shape of the edges structure. In fact, it is based on the structure tensor that represents a very interesting concept for characterizing the local shape. On the other hand, the use of the histogram concept makes the proposed descriptor so effective and useful when a reduced feature representation is required. In this paper, the proposed HoST descriptor is addressed to the pattern clustering task. An extensive experimental validation demonstrates its performance when compared to other existing feature descriptors such as Local Binary Patterns and Histogram of Oriented Gradients. In addition, the proposed descriptor succeeds in improving the performance of clustering based resolution enhancement approaches.