使用GIoU改进非极大值抑制的目标检测算法

Object Detection Algorithm for Improving Non-Maximum Suppression Using GIoU

针对单阈值-非极大值抑制算法中出现的目标漏检和重复检测问题,本文提出了一种使用全局交并比指标GIoU(Generalized Intersection over Union)衡量目标相似度的双阈值非极大值抑制算法GDT-NMS(Generalized Dual Threshold NMS,GDT-NMS).使用双阈值改进NMS算法和soft-NMS算法,抑制多余的检测框;在此基础上,使用GIoU替换传统的IoU计算目标间的相似度,使目标的定位更加准确;进一步,使用非线性函数赋予检测框不同比例的权值惩罚,使检测框的得分随距离呈非线性变化,目标区分度更高.改进算法在PASCAL VOC和MSCOCO上的检测精度分别为74.8%和25.9%,与使用NMS算法作为后处理的Faster R-CNN算法相比,性能分别提升了1.6%和1.5%.同时本文算法具有较快的检测速度.

Aiming at the problem of missed detection and repeated detection in the single-threshold-non-maximum suppression algorithm,this paper proposes a dual-threshold Non-Maximum Suppression algorithm using GIoU(Generalized Intersection over Union).Using dual thresholds to improve the NMS algorithm and the soft-NMS algorithm,suppressing redundant detection boxes,not only balances the relationship between the object missed detection problem and the object false detection problem caused by the single threshold algorithm,but also reduces the occurrence of the soft-NMS algorithm.Based on the above,using GIoU instead of IoU to calculate the similarity between objects,so that the positioning of the object is more accurate;the non-linear function is used to give different weights to the proposal boxes,which makes the proposal boxes’scores change non-linearly with distance,and the target discrimination is higher,which is more conducive to suppressing the proposal boxes.The detection accuracy of the improved algorithm on PASCAL VOC and MSCOCO is 74.8%and 25.9%,respectively.At the same time,the algorithm in this paper has a fast detection speed.