基于混合池化YOLO的目标检测方法

Object Detection Method Based on Mixed-pooling YOLO

针对目标检测方法中存在的正负样本分布不均衡、检测精度低等问题,提出了一种基于混合池化YOLO的目标检测方法(object detection method based on mixed-pooling YOLO,ODMMP-YOLO).ODMMP-YOLO首先将混合池化融入Darknet-53网络构造出一种新颖的DMP(darknet based on mixed pooling)网络模型,能够在下采样阶段有效减少特征图信息的丢失,从而提升识别精度;然后采用GIoU(generalized intersection over union)定位损失衡量真实边框与预测边框之间的距离,有效地提升定位精度;在计算置信度损失时给予误分检测框更多的loss惩罚,有效解决正负样本分布不均衡的问题.在PASCAL VOC 2007数据集上对ODMMP-YOLO进行验证,实验结果表明:与传统YOLOv3目标检测方法相比,ODMMP-YOLO识别部分单独类别时的平均精度AP提升约15%,在识别所有类别时的均值平均精度mAP值提升约5%;与其他主流检测方法相比,ODMMP-YOLO能够更好地识别与定位日常生活中的常见目标物体,且具有较好的视觉效果.

Aiming at the problems of unbalanced sample distribution and low detection accuracy in object detection methods,the authors proposed an object detection method based on mixed-pooling YOLO(ODMMP-YOLO).Firstly,ODMMP-YOLO constructed a new type of DMP(Darknet based on Mixed Pooling)feature extraction network by integrating the mixed pooling into the Darknet-53 network,which effectively reduced the information loss of feature map in the down-sampling stage,thereby improving the accuracy of recognition.Secondly,the localization loss of Generalized Intersection over Union(GIoU)was used to measure the distance between the true bounding box and the predicted bounding box,which efficiently improved the localization accuracy.Finally,when calculating the confidence loss,more loss penalties were given to the misclassification detection bounding box,which effectively solved the problem of the unbalanced sample distribution.ODMMP-YOLO is verified on PASCAL VOC 2007 Dataset,and the experimental results show that compared with the traditional YOLOv3 object detection method,the average precision of the ODMMP-YOLO in identifying some separate classes is improved by nearly 15%.When identifying all classes,the ODMMP-YOLO increases the mean average precision mAP by about 5%.Compared with some mainstream detection methods,ODMMP-YOLO can better identify and locate common objects in daily life,and has a better visual effect.