Texture-based rotation-invariant Histograms of Oriented Gradients

Microorganism detection through computer vision is of great essence in many scientific and industrial fields, among which vorticella is always a challenging and focused topic, especially for the sake of water quality monitoring and biological diversity tracking. However, given the fact that vorticella is likely to be obscured by alga and it float in different poses, the visual detecting method must be tolerant of different rotations and slight ambiguity. Previously, the Histograms of Oriented Gradients (HOG) is widely used in pedestrian, hand gesture and many other object detections. And the Local Binary Pattern (LBP) has been proven to be a widely applicable image feature for texture classification and face analysis. By combining the two methods, this paper presents a method to build rotation-invariant descriptors, in which the idea of dividing whole image into spatial blocks to obtain local relative distribution and the concept of utilizing the discrete Fourier transform (DFT) on cycle-shift patterns in histogram to acquire rotation-invariant features are efficiently bonded.