Anti-rotation Binary Descriptor for Circular Image Recognition

Based on pairs of spatial symmetric patches, a novel efficient and distinctive binary descriptor was proposed in this paper. To achieve rotation invariance during feature computation, a local coordinate system was found with radial transform technology. On the basis of that, local binary patterns against rotation could be extracted. Meanwhile, the circular image was divided into a set of overlapped annular regions, and pairs of patch description were constructed with histogram within the ring. Finally the rotation invariant image description could be obtained by concatenating the ring features from inner to outer. The performance of proposed method was tested with 3 data sets. The test results showed that its recognition accuracy reached 100%, 100% and 97.07% respectively, which were superior to the results of traditional methods based on LBP feature. Moreover, the algorithm was efficient, as single point feature computation needed only 0.045ms.