首页|基于改进YOLOv4-Tiny的机械零件目标检测算法

基于改进YOLOv4-Tiny的机械零件目标检测算法

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
随着智能化生产的普及,机械零件的智能装配技术受到广泛关注,为了改善传统特征提取算法中检测误差大、精度低等问题,以常见机械零件为研究目标,结合深度学习算法中的轻量级网络为基础模型进行优化,使用YOLOv4-Tiny中的CSP-Darknet53 作为特征提取网络,在特征提取网络后添加改进后的MA-RFB模块,引入多分支卷积和空洞卷积加强感受野.并对颈部网络进行改进,选择PANet代替FPN,并添加注意力模块CBAM,形成CM-PANet对零件目标进行多尺度检测,在自制的零件数据集AP达到 96.47％,检测速度达到 0.001 38 s每样本.相比于原版YOLOv4-Tiny网络AP提高了2.80％,改进后的算法在速度和精度达到了一个平衡,体现了研究的理论和应用价值.
Recognition of Mechanical Parts Based on Improved YOLOv4-Tiny Algorithm
In order to improve the problems of large detection error and low accuracy in traditional mechanical parts feature extraction algorithm,common mechanical parts are taken as the research target and lightweight network in deep learning algorithm is adopted as the base model for optimization.CSP-Darknet53 is used to extract the feature.An improved MA-RFB module is added after the feature extraction network,and multi-branch convolution and empty convolution was introduced to strengthen the receptive field.In addition,the neck network is improved,PANet is selected to replace FPN,and the attention module of CBAM is added to form RC-PANet for multi-scale detection of parts targets.AP reaches 96.47％in the self-made part dataset,and the detection speed is 0.001 38 s per sample.Without losing too much speed,compared with the original YOLOv4-Tiny network,AP improves by 2.80％,and the improved algorithm achieves a balance in speed and precision,which reflects the theoretical and application value of the research.

image classificationlightweight networkdeep learningmechanical partsreceptive field module

杜保帅、房桐、赵燕成、赵景波

展开 >

青岛理工大学信息与控制工程学院,山东 青岛 266520

图像分类 轻量级网络 深度学习 机械零件 感受野

国家自然科学基金项目山东省自然科学基金项目山东省高等学校科技计划项目青岛市科技计划项目青岛市民生计划项目

51475251ZR2013FM014J12LN3715-9-2-109-nsh22-3-7-xdny-18-nsh

2024

10.3969/j.issn.1004-1699.2024.02.015

传感技术学报
东南大学 中国微米纳米技术学会

传感技术学报

CSTPCD北大核心
影响因子：1.276
ISSN：1004-1699
年,卷(期)：2024.37(2)
  • 18