Histograms of dominant orientations (HDO) is a simple local image descriptor with fine performance.However,the original HDO feature description has no rotation invariance.This paper presents a rotation-invariant HDO feature description.To acquire the rotation invariant feature,i.e.,dominant orientation and the coherent,by RGT (radial gradient transform),the structure tensor of given location is calculated in a circular neighborhood.Then,to enhance distinctiveness,space pooling operation is implemented with multi-sector division.Test results in public MIT faces data show that if the image does not rotate,the proposed method and the original HDO descriptor almost have the same accuracy (92.10 ％),while,if the image rotates,the accuracy of the improved HDO descriptor is higher than that of the original HDO by 10.36％.In addition,in the experiments of pedestrians,synthetic rotated palms and faces detections,our method is obviously superior to its original one.Moreover,the proposed method shows better recognition accuracy than most recent anti-rotation descriptors in public 53Objects,ZuBuD and Kentuky image datasets.%主方向直方图(Histograms of dominant orientations,HDO)是一种简单但性能优良的局部图像描述子,但是,原有的HDO特征描述不具备旋转不变性.本文提出一种抗旋转变换HDO特征描述方法,在进行RGT (Radial gradient transform)变换后,采用圆形邻域计算给定位置的结构张量,使得求取的主方向和一致性特征分量具备一定的旋转不变性,最后为增强辨别能力,采用了多扇区划分空间池化操作.在公开的MIT人脸数据集中的测试结果显示,如果图片不旋转,本文方法准确率与传统的HDO算法基本持平,达到92.10％,但当样本图片旋转后,本文算法准确率比传统HDO算法高10.36％.此外,在行人数据集、合成的旋转手掌和旋转人脸识别实验中,本文方法的检测结果也明显优于传统的HDO算法.另外本文方法在53Objects、ZuBuD和Kentuky三个数据集上的识别性能也优于大部分现有抗旋转算子.
展开▼
