基于热扩散理论的窗融合方法研究

Detection Windows Fusion Based on Theory of Heat Diffusion

窗融合是滑动窗目标检测方法中的一个重要步骤。针对传统方法的缺陷,提出了一种新的窗融合方法。该方法把每个初始窗口当作系统中的一个位置,两个窗口的检测分数和重叠面积用来计算对应位置之间的热传导系数,最终利用线性各向异性热扩散条件下系统温度之和最大化问题来模拟窗融合工作。采用贪婪算法获得目标函数的近似最优解,相应的热源即为窗融合结果。在VOC2009目标数据库和INRIA行人数据库上的实验显示,该方法不仅能够删除重复检测,还可以排除误检以及防止相邻目标干扰。相比传统的非极大值抑制方法,该方法在不损失召回率的前提下显著地提升了目标的检测精度。

Detection windows fusion is an important step of object detection based on sliding window. To overcome shortcomings of traditional detection fusion methods, this paper proposes a novel one. The method treats every preliminary window as a location in system, and heat conductivity between two locations is calculated by detection scores and overlapping area of corresponding windows. Finally, the detection windows fusion task is modeled by temperature maximization on linear anisotropic heat diffusion, of which the temperature maximization with finite K heat sources corresponds to K final windows. This paper obtains a near-optimal solution of objective function by a greedy algorithm. Experimental results on VOC2009 and INRIA pedestrian datasets show that our method not only deletes overlapping detections, but also rejects false positives and prevents interference between adjacent objects. Compared with traditional non maximum suppression, our method can obtain higher detection precision without loss of recall rates.