基于改进YOLOv8的细长物体检测方法
Method of Detecting Slender Object Based on Improved YOLOv8
李佳兴 1文峰1
作者信息
- 1. 沈阳理工大学信息科学与工程学院,沈阳 110159
- 折叠
摘要
在计算机视觉领域,无锚框的检测算法能够较好地检测并定位任意几何形状的物体,有利于检测细长物体.从更好保留细长物体信息、防止信息的混淆以及更好表达非局部信息之间关联的角度出发,提出改进YOLOv8的细长物体检测方法.在YOLOv8的骨干和头部网络中加入有助于处理细长物体的SPD-Conv构建块,SPD-Conv充分利用空间分割与非跨步卷积处理技术,有效减少细长物体特征的丢失;改进YOLOv8骨干网络中的卷积运算,解决参数共享的问题,对图像不同位置和不同通道的特征赋予不同的重要性和含义;为更好标定细长目标,使用非局部non_local注意力机制理解图像中的长距离依赖关系和相关性.实验结果表明:与YOLOv8原型相比,改进算法平均精度提升了 11.35％,推理速度略有降低,为每秒20帧,基本达到了实时性要求.
Abstract
In the field of computer vision,the detection algorithm without anchor frame can be used for detecting slender objects and better locate objects with any geometric shape.From the perspec-tive of better retaining the information of elongated objects,preventing the confusion of information and better expressing the correlation between non-local information,an improved elongated object detection method of YOLOv8 is proposed.In the backbone and head network of YOLOv8,SPD-Conv building blocks are added to help process elongated objects,and SPD-Conv fully uses spatial segmentation and non-step-step convolution processing techniques to effectively reduce the loss of features of elongated objects.The convolution operation in the YOLOv8 backbone network is im-proved,the problem of parameter sharing is solved,and different importance and meaning are atta-ched to the features of different positions and channels of the image.In order to better calibrate slender targets,attention mechanism non_local is used to learn long-distance dependencies and cor-relations in images.Experimental results show that the average accuracy of the improved algorithm is increased by 11.35％compared with the YOLOv8 prototype,and although the inference speed is slightly reduced,the inference speed of 20 frame per second is maintained,which basically meets the real-time requirements.
关键词
YOLOv8/目标检测/注意力机制/细长物体识别Key words
YOLOv8/object detection/attention mechanism/slender object recognition引用本文复制引用
基金项目
国家重点研发计划"社会治理与智慧社会科技支撑"重点专项(2022YFC3302502)
出版年
2024
NETL
NSTL
万方数据