基于主方向的旋转不变HOG特征

Rotation-invariant HOG feature based on main orientation

特征提取是目标检测与识别领域的研究热点之一,HOG(Histogram of Oriented Gradient)特征由于其对图像局部信息良好的几何和光照不变性,在行人检测、车牌和人脸识别等计算机视觉邻域得到了广泛应用,但是HOG不具有旋转不变的特性,使得该特征在实际应用中存在着一些局限性。针对该问题,提出一种具有旋转不变性的HOG特征提取方法,首先根据图像梯度信息提取主方向并设置为参考方向,接着旋转主方向至参考方向,在旋转后的图像上得到旋转不变的HOG特征。并且设计了一种面向图像匹配的相似性度量准则,它以单个图像块(Block)特征向量为基元,与待匹配图像中对应块及其邻域块特征向量的相似度共同作为度量标准,增强了旋转图像在像素平移情况下的匹配效果。实验结果表明,提出的改进HOG特征具有良好的旋转不变特性。

abstract： the feature extraction is one of the research hotspots in the field of target detection and recognition. HOG （histo- gram of oriented gradient） feature has been wi.dely used in the field of computer vision, such as pedestrian detection, license plate recognition and face identification due to its geometrical and illumination invariability for local image. However, HOG fea- ture has limitation in practical application because it has no rotation-invariant characteristic. Therefore, an extraction method of rotation-invariant HOG feature is proposed in this paper. Firstly, the main orientation is extracted according to the image gra- dient information and is set as reference direction, then the rotation-invariant HOG features are obtained from the rotated image by turning the main direction to the reference direction. A similarity measurement criterion for image-oriented matching is de- signed in this paper, in which the single image block feature vector is took as element, and is used to be the measurement stan- dard with the similarity of the block corresponding to the image under matching and its adjacent block feature factors. The experimental results show that the improved HOG feature has good rotational invariance.