多方向线积分的梯度特征

典型的梯度特征包括HOG(梯度方向直方图)、Shapelet及Edgelet等,这些特征被广泛用于目标检测、目标识别、图像检索及场景分类等领域。针对HOG特征运算复杂度高的问题,提出了一种新的多方向线积分的梯度特征(MDIG)。通过避免计算梯度方向并利用积分图,该特征简化了计算过程,提高了计算速度,因而便于在DSP等硬件上实现。实验中新特征被应用于人体头肩检测。实验结果表明当使用AdaBoost算法训练分类器时,该特征的描述能力与HOG相当,同时其计算时间仅为HOG的1/3,整体性能优于HOG。最后,针对梯度特征的适用范围对其未来应用的发展方向进行了讨论。

Multi-directional integration of gradients

Representative algorithms for extracting gradient features include HOG (histogram of oriented gradients), Shapelet and Edgelet. They have been widely used in the fields of object detection, object recognition, image retrieval,and scene classification. A new feature named Multi-Directional Integration of Gradients (MDIG) is proposed after analyzing the HOG and its fast version. By avoiding computing the gradient orientation and using an integral image, the new feature is easier to extract and can be hardware implemented on a DSP.The MDIG is tested on images of human head-shoulders. Using AdaBoost to train a classifier, the result is comparable with HOG, while the computation burden is only one third of that of the HOG. Finally, the use of gradient feature is discussed, and the future work is summarized.