计算机工程 ›› 2019, Vol. 45 ›› Issue (11): 243-248.doi: 10.19678/j.issn.1000-3428.0052861

• 图形图像处理 • 上一篇    下一篇

嵌入SENet结构的改进YOLOV3目标识别算法

刘学平1, 李玙乾1,2, 刘励1,2, 王哲1,2, 刘宇3   

  1. 1. 清华大学 深圳研究生院, 广东 深圳 518055;
    2. 清华大学 机械工程系, 北京 100084;
    3. 长虹智能制造技术有限公司, 成都 621000
  • 收稿日期:2018-10-12 修回日期:2018-11-23 发布日期:2018-12-04
  • 作者简介:刘学平(1965-),男,副教授、博士,主研方向为目标识别、智能控制;李玙乾、刘励、王哲,硕士研究生;刘宇,工程师。
  • 基金项目:
    国家自然科学基金(51475263)。

Improved YOLOV3 Target Recognition Algorithm with Embedded SENet Structure

LIU Xueping1, LI Yuqian1,2, LIU Li1,2, WANG Zhe1,2, LIU Yu3   

  1. 1. Graduate School at Shenzhen, Tsinghua University, Shenzhen, Guangdong 518055, China;
    2. Department of Mechanical Engineering, Tsinghua University, Beijing 100084, China;
    3. Changhong Intelligent Manufacturing Co., Ltd., Chengdu 621000, China
  • Received:2018-10-12 Revised:2018-11-23 Published:2018-12-04

摘要: 为准确识别工业图像中的目标零件,提出一种改进的YOLOV3目标识别算法。结合K-means聚类与粒子群优化算法进行锚框计算,以降低初始点对聚类结果的影响,加快算法收敛速度。同时在YOLOV3网络shortcut层嵌入SENet结构,得到SE-YOLOV3网络。对零件图像进行数据增强并加入零件标注,制作包含10 816张图片的样本集,用于算法训练和测试。实验结果表明,该算法能够获得平均交并比为83.01%的锚框,当样本图像存在较多残缺零件干扰时,YOLOV3存在将背景识别为零件的情况,其查准率与查全率分别为72.11%和97.51%,而SE-YOLOV3能有效减少假正例数量,其查准率与查全率分别为90.39%和93.25%。

关键词: 目标识别, 卷积神经网络, SENet结构, YOLOV3网络, 粒子群优化算法

Abstract: In order to accurately identify the target parts in the industrial image,an improved YOLOV3 target recognition algorithm is proposed.The K-means clustering and particle swarm optimization algorithm are combined to calculate the anchor box to reduce the influence of the initial point on the clustering result and speed up the convergence of the algorithm.The SE-YOLOV3 network is obtained by embedding the SENet structure after the shortcut layers.A sample set containing 10 816 images is created by collecting the part images and enhancing the data while labeling the part in the image,which is used to train and test the network.Experimental results show that the proposed algorithm can obtain an anchor box with an average IoU of 83.01%.When there are more defective parts in the sample image,YOLOV3 might identify the background as a part,and the precision and recall rate are 72.11% and 97.51% respectively,while the SE-YOLOV3 can accurately identify target parts,whose precision and recall rate are 90.39% and 93.25% respectively.

Key words: target recognition, Convolutional Neural Network(CNN), SENet structure, YOLOV3 network, Particle Swarm Optimization(PSO) algorithm

中图分类号: