Orientation and scale invariant binary descriptor based on Haar wavelet

Abstract

In this paper, an orientation and scale invariant binary descriptor is proposed, which can be used in key-points matching systems. Conventionally, a binary descriptor is generated by comparing the intensities of pixels directly, such as those in Binary Robust Independent Elementary Features (BRIEF) and Oriented FAST and Rotated BRIEF (ORB). However, comparing intensities of pixels may lose the texture information in the region of interest, and lead to a high false match rate in a practical application. In our proposed method, the region of interest is segmented into grid cells and then the binary Haar wavelet responses are computed to store the texture information of the patch. Concretely, the texture information in each cell is expressed by the horizontal and vertical gradients and the polarity of intensity changes which are indicated by four components of Haar wavelet response. The binary descriptor is generated by comparing the Haar wavelet response in each pair of grid cells. Furthermore, to be scale and orientation invariant, the patch of key-points is rotated to the primary direction of the centroid vector in the image pyramid. Extensive experimental results show that our descriptor significantly outperforms other five state-of-the-art binary descriptors in key-point matching systems. The average percentage of correct matches of our method is 32.79% higher than that for FREAK and 5.31% higher than that for LDB.