摘要
针对传统目标检测方法不能兼顾目标识别精度和检测实时性,且在实际生产复杂工况下识别效果不佳的问题,提出一种基于Inception-SSD框架的零件深度学习识别方法。首先,提出了融合Inception预测结构的SSD优化框架Inception-SSD,将Inception网络结构引入到SSD网络额外层中,并使用批量标准化模块(BN)和残差结构连接,从而捕获更多目标信息而又不会增加网络复杂性,以提高检测准确率而又不影响其检测速度,并增加算法鲁棒性;然后提出在原损失函数基础上增加排斥损失项以改进损失函数,同时采用一种基于加权算法的非极大值抑制方法,克服模型表达能力不足的缺点。最后,将改进前后SSD算法在自制零件数据集上进行训练和测试,实验结果表明本文方法在实际生产过程复杂情况下检测准确率达到97.8%,相比原SSD算法提升11.7%,检测速率41 frame/s。在提高检测精度同时还保证了实时性,能够满足实际生产环境零件检测需求。
In traditional target detection methods,there is a trade-off that exists between target detection accuracy and real-time detection,and the recognition accuracy is inferior under actual,complex production scenarios.To address this,a deep learning detection method based on the Inception-SSD framework was herein proposed.In this framework,an inception network structure was introduced into the extra layer of the SSD network,and batch normalization(BN)and residual structure connection were used to capture target information without increasing network complexity.Owing to this,detection accuracy was improved without the real-time detection performance being affected and the algorithm also becomes more robust.Subsequently,the exclusion loss term based on the original loss function increases,which in turn improves the loss function.Furthermore,a non-maximum suppression weighting method was used to overcome the shortcomings of insufficient expression ability of the model.Finally,the improved SSD algorithm was trained and tested on a self-made dataset and compared with the original and the latest inception-SSD algorithms.Experimental results show that the detection accuracy of the proposed method is 97.8%in an actual production process,which is an improvement of 11.7 percentage points over the original SSD algorithm,and the detection speed is 41 fps.Therefore,the proposed method exhibits superior real-time performance,thereby meeting actual production demands.
作者
余永维
韩鑫
杜柳青
YU Yong-wei;HAN Xin;DU Liu-qing(College of Mechanical Engineering, Chongqing University of Technology, Chongqing 400054, China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2020年第8期1799-1809,共11页
Optics and Precision Engineering
基金
重庆市基础与前沿研究计划基金资助项目(No.cstc2017jcyjAX0344)
国家自然科学基金资助项目(No.51775074)。