目标鲁棒识别的抗旋转HDO局部特征描述

发布时间：2019-04-17 13:06

【摘要】：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.
[Abstract]:The principal direction histogram (Histograms of dominant orientations,HDO) is a simple but excellent local image descriptor, but the original HDO feature description does not have rotation invariance. In this paper, an anti-rotation transform HDO feature description method is proposed. After the RGT (Radial gradient transform) transformation, the circular neighborhood is used to calculate the structure Zhang Liang at a given position, so that the obtained principal direction and consistent feature components have certain rotation invariance. Finally, in order to enhance the discrimination ability, multi-sector partition spatial pooling operation is adopted. The test results in the public MIT face dataset show that if the image is not rotated, the accuracy of this method is basically the same as that of the traditional HDO algorithm, reaching 92.10%, but when the sample image is rotated, The accuracy of the proposed algorithm is 10.36% higher than that of the traditional HDO algorithm. In addition, in pedestrian data sets, synthetic rotating palm and rotating face recognition experiments, the detection results of this method are obviously better than the traditional HDO algorithm. In addition, the recognition performance of the proposed method on 53 objects, ZuBuD and Kentuky data sets is better than that of most of the existing anti-rotation operators.
【作者单位】：湘潭大学信息工程学院;机器人视觉感知与控制技术国家工程实验室;
【基金】：湖南省自然科学基金(2017JJ2251)资助~~
【分类号】：TP391.41

【相似文献】