基于梯度双面互补特性的级联快速目标检测

发布时间：2018-06-08 23:02

本文选题：目标检测 + 类无关和类相关　；参考：《电子学报》2017年10期

【摘要】：针对目标检测中精度和速度难以兼顾的问题,借助视觉注意理论中的目标感知与识别机制,分析目标描述中梯度幅值与梯度方向信息之间具有的互补性,提出了基于两层级联梯度特征的快速目标检测模型,可有效描述类无关和类相关检测器.一方面,采用梯度幅值特征,从滑动窗口采样中获得候选目标提议,大幅降低了验证窗口的数量,确保检测速度,另一方面,利用级联方式学习训练多个子检测器,可更好实现不同尺度变化下的目标检测精度.PASCAL数据集上的实验结果,解释了级联梯度特征对目标结构描述的有效性,表明了该文方法在与现有先进方法的检测精度相当的前提下,可极大提升检测速度.
[Abstract]:In view of the problem that the precision and speed of the target detection are difficult to take into account, the complementarity between the gradient and the gradient direction information in the target description is analyzed with the aid of the target perception and recognition mechanism in the visual attention theory. A fast target detection model based on the cascaded gradient features of the two layer is proposed, which can effectively describe the class independence and the class correlation. On the one hand, the gradient amplitude feature is used to obtain the candidate target proposal from the sliding window sampling, which greatly reduces the number of the verification windows and ensures the detection speed. On the other hand, the cascade mode is used to train the multiple sub detectors, and the experiment on the.PASCAL data set of the target detection precision under different scales can be better realized. As a result, the effectiveness of cascaded gradient features to the description of the target structure is explained. It shows that the proposed method can greatly improve the detection speed on the premise that the detection precision of the existing advanced methods is equal.
【作者单位】：合肥工业大学计算机与信息学院;
【基金】：国家自然科学基金(No.61273237,No.61503111)
【分类号】：TP391.41

【相似文献】