摘要
原始YOLOv3模型被认为适合求解多尺寸的图像目标检测问题,但是对于粮虫小目标检测存在表征能力不足且检测效率较低的问题。本文基于此提出了一种融合GIoU算法的YOLOv3检测模型,一方面使用GIoU算法弥补IoU算法对于两个不相交box无法进行量化的问题,同时使用GIoU对损失函数进行优化,损失函数优化为GIoU损失、置信度损失和分类损失三方面;另一方面使用五种数据增强手段对原始1998张数据集进行数据增强,最终形成大小为9990张的数据集,并使用K-means聚类算法对自制数据集进行聚类分析,聚类出符合粮虫小目标检测的先验框。针对自制的9990张粮虫的数据集进行实验获得了99.43%的mAP和每幅图像0.040 s的检测速度,与原始YOLOv3模型相比,本文所提模型对于小目标的粮虫检测效果得到了很大的提升。
The original YOLOv3 model was considered to be suitable for solving multi-size image object detection problem,but there were problems such as insufficient characterization ability and low detection efficiency for grain pest small object detection.Based on this,this paper proposes a detection model YOLOv3 that integrates GIoU algorithm.On the one hand,GIoU algorithm is used to make up for the problem that IoU algorithm cannot quantify two non-overlapping boxes.At the same time,GIoU is used to optimize the loss function,which is optimized for GIoU loss,confidence loss and classification loss.On the other hand,five data enhancement methods are used to enhance the original 1998 dataset,and 9990 dataset are finally formed.And k-means clustering algorithm is used to conduct clustering analysis on the self-made dataset,and a set of prior bounding boxes suitable for grain pest small object detection are clustered.Based on the dataset of 9990 grain pests,optimized model achieves a mAP 99.43%with the detection speed of 0.040 s peer image.Compared with the original YOLOv3 model,the detection effect of the proposed model for small targets has been greatly improved.
作者
吕宗旺
金会芳
甄彤
孙福艳
桂崇文
LüZongwang;Jin Huifang;Zhen Tong;Sun Fuyan;Gui Chongwen(College of Information Science and Engineering,Henan University of Technology,Zhengzhou 450001;Key Laboratory of Grain Information Processing and Control,Ministry of Education,Zhengzhou 450001)
出处
《中国粮油学报》
CAS
CSCD
北大核心
2021年第10期159-165,共7页
Journal of the Chinese Cereals and Oils Association
基金
国家重点研发计划(2017YFD0401004)。