摘要
为了解决自然场景包裹破损检测中由于目标形态与尺度多样、模型耗时过长造成的检测难题,设计了一种基于通道注意力机制与快速空间金字塔池化(Space Pyramid Pool-Fast,SPPF)的轻量级包裹破损检测算法。在YOLOv5s的基础上,使用改进的ShuffleNetV2网络模型对其主干结构进行轻量级优化,降低模型计算量,提高检测速度;在模型的主干网络部分引入通道注意力机制——Squeeze Excitation(SE),减少了卷积神经网络对图像相关特征的重复提取,提高信息的表征能力;利用SPPF有效避免了对图像区域裁剪、缩放操作导致的图像失真,有效减少误检与漏检。在数据集上的测试结果表明,该方法对包裹图像的检测速度达到了68.5帧/秒,模型计算量仅为2.5 GFLOPs,与YOLOv5s相比,检测速度提升了105.7%,模型计算量下降了84.2%,利于边缘计算设备部署。
In order to solve the detection problem of package damage in natural scenes due to the diversity of target shapes and scales,as well as the long model time,a lightweight package damage detection algorithm based on channel attention mechanism and Space Pyramid Pool-Fast(SPPF)is designed.On the basis of YOLOv5s,an improved ShuffleNetV2 network model is used to perform lightweight optimization for its backbone structure,so as to reduce model computation and improve detection speed;the channel attention mechanism Squeeze Exception(SE)is introduced in the backbone network of the model to reduce the repeated extraction of image related features by convolutional neural networks and improve the information representation ability;the use of SPPF effectively avoids image distortion caused by clipping and scaling operations on image regions,effectively reducing false positives and missed detections.The test results on the data set show that the detection speed of the proposed method for the package image reaches 68.5 frame/s,and the calculation amount of the model is only 2.5 GFLOPs.Compared with YOLOv5s,the detection speed is increased by 105.7%,and the calculation amount of the model is reduced by 84.2%,which is conducive to the deployment of edge computing devices.
作者
周耀威
孔令军
李慧刚
郭乐婷
杨文杰
陈一品
张栋濠
ZHOU Yaowei;KONG Lingjun;LI Huigang;GUO Leting;YANG Wenjie;CHEN Yipin;ZHANG Donghao(School of Networking and Communication Engineering,Jinling Institute of Technology,Nanjing 211169,China;Research and Social Services Division,Zhejiang Tourism and Health College,Zhoushan 316111,China)
出处
《无线电工程》
北大核心
2023年第11期2626-2634,共9页
Radio Engineering
基金
江苏省大学生创新训练项目(202213573025Z)
江苏省高等学校基础科学(自然科学)研究重大项目:集成计算的固态存储系统中高可靠性存储技术研究(22KJA510009)
金陵科技学院高层次人才科研启动资金(jit-b-202110)。
关键词
深度学习
包裹破损检测
图像处理
轻量化
注意力机制
deep learning
package damage detection
image processing
lightweight
attention mechanism
