description for 3d images with histogram of gradients based on maximum directional derivative

The histogram of gradients for three-dimensional images (HOG3D) has been widely used to generate descriptors of local features for the matching of salient interest blobs between volumetric medical images. The distinctiveness of descriptors is essential to match similar patterns as well as differentiate different patterns. However, most methods for HOG3D are based on the unequal azimuth-elevation-angle division of the three-dimensional orientations, leading to the reduction of the distinctiveness. This paper presents a description method to extract distinctive descriptors by using a novel histogram of gradients based on the maximum directional derivative (HOG3DMAX). The three-dimensional orientations are divided equally into 24 homogeneous regions for the bins of gradient histogram to improve the distinctiveness. In addition, the interpolation of gradients is applied in HOG3DMAX according to the relatively simple form of the homogeneous regions in order to avoid harmful effects from boundary in images. Experiments were performed on three sets with different extents of similarity. The cluster analysis and classification accuracy were utilized to validate the proposed method. Compared with the commonly used azimuth-elevation based method, the new method shows lower intra-cluster pairwise distances relative to inter-cluster pairwise distances, and achieved 5% higher classification accuracy. The results indicate the improvement of distinctiveness by using equally divided regions and the interpolation of gradients. We conclude that the HOG3DMAX is an effective description method for three-dimensional images.The histogram of gradients for three-dimensional images (HOG3D) has been widely used to generate descriptors of local features for the matching of salient interest blobs between volumetric medical images. The distinctiveness of descriptors is essential to match similar patterns as well as differentiate different patterns. However, most methods for HOG3D are based on the unequal azimuth-elevation-angle division of the three-dimensional orientations, leading to the reduction of the distinctiveness. This paper presents a description method to extract distinctive descriptors by using a novel histogram of gradients based on the maximum directional derivative (HOG3DMAX). The three-dimensional orientations are divided equally into 24 homogeneous regions for the bins of gradient histogram to improve the distinctiveness. In addition, the interpolation of gradients is applied in HOG3DMAX according to the relatively simple form of the homogeneous regions in order to avoid harmful effects from boundary in images. Experiments were performed on three sets with different extents of similarity. The cluster analysis and classification accuracy were utilized to validate the proposed method. Compared with the commonly used azimuth-elevation based method, the new method shows lower intra-cluster pairwise distances relative to inter-cluster pairwise distances, and achieved 5% higher classification accuracy. The results indicate the improvement of distinctiveness by using equally divided regions and the interpolation of gradients. We conclude that the HOG3DMAX is an effective description method for three-dimensional images