改进的YOLOv4复杂构件分类识别算法  

Classification and recognition algorithm for complex components based on improved YOLOv4

作　　者：林鑫 沈建新[1] 秦顺 潘峰 LIN Xin;SHEN Jian-xin;QIN Shun;PAN Feng(College of Mechanical and Electrical Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)

机构地区：[1]南京航空航天大学机电学院,江苏南京210016

出　　处：《计算机工程与设计》2022年第9期2519-2524,共6页Computer Engineering and Design

基　　金：国防重大项目培育基金项目(NP2020420)。

摘　　要：为解决航天设备复杂构件分类检测时缺少数据集、识别困难等问题,在三维建模软件上对构件不同角度截图制作数据集,对数据进行增广。提出一种基于改进YOLOv4算法的航天设备复杂构件分类识别算法,将YOLOv4算法中残差卷积块的个数由5个调整为3个,各部分通道数缩减一半。只采用上采样结构,省略下采样结构,减少训练参数,精简网络结构。实验结果表明,改进的YOLOv4目标检测算法训练速度快,检测精度高,在航天复杂构件图像验证集上的误差降至0.66,mAP达到97.63%,性能优于同类算法。To solve the problems of lack of datasets and difficulty in identification and detection of complex components in aerospace equipment,screenshots of the components were took from various angles in 3D modeling software to make the datasets,and the datasets were augmented.A classification and detection algorithm of aerospace complex components based on improved YOLOv4algorithm was proposed.The number of residual convolution blocks was adjusted from 5to 3,and the number of channels of various parts was reduced by half.Only the up-sampling structure was used and the down-sampling structure was omitted,which could reduce a large number of training parameters and streamline the network structure.Experimental results show that the improved YOLOv4target detection algorithm has high training speed and detection accuracy.The error on the complex components of aerospace verification datasets is reduced to 0.66and mAP reaches 97.63%.The performance of this algorithm is better than that of similar algorithms.

关 键 词：航天设备 复杂构件 数据增强 深度学习 识别分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]