基于Transformer改进的Faster-Rcnn仓储箱体检测算法  

Storage Box Detection Method Based on Transformer Improved Faster-Rcnn

作　　者：李映松 杨爱英[1,2] 刘轩 畅宇堃 LI Yingsong;YANG Aiying;LIU Xuanxuan;CHANG Yukun(School of optoelectronics,Beijing University of technology,Beijing 100081,China;Key Laboratory of information photonic technology,Ministry of industry and information technology(Beijing University of Technology),Beijing 100081,China)

机构地区：[1]北京理工大学光电学院,北京100081 [2]信息光子技术工信部重点实验室(北京理工大学),北京100081

出　　处：《自动化与仪器仪表》2022年第8期1-6,共6页Automation & Instrumentation

基　　金：国家自然科学基金资助项目(61971046)。

摘　　要：为解决传统目标检测方法准确性差、效率低,无法满足智能仓储场景需求的问题,提出基于Transformer改进的Faster-Rcnn仓储箱体检测模型。首先,在Faster-Rcnn模型的基础上,将卷积神经网络Resnet50改进为Swin Transformer模型,使用Swin Transformer进行全局信息提取,解决了使用传统算法特征提取不理想,产生冗余的检测窗口以及误检窗口的问题。其次,引入了特征金字塔结构,使模型适用于多尺度的物体检测。最后,使用ROI Align代替ROI Pooling,消除了ROI Pooling中因浮点数取整从而对后层的检测框回归产生的误差。在自建的仓储数据集训练模型,将数据集图片进行随机旋转、随机剪裁、图片标准化等操作进行数据增强。实验结果表明,改进后的模型用于箱体检测,平均准确率达到90.6%,平均召回率达到93.3%,平均检测速度达到8.9fps,较好地实现了仓储物体的准确检测,满足智能仓储的需求。准确率方面比Faster R-CNN、YOLOv3、SSD、FCOS等算法高出6.1%、5%、10.2%、9.7%,召回率高出了5.9%、4%、10.1%、9.4%。In order to solve the problems of poor accuracy and low efficiency with traditional object detection methods which couldn’t satisfy the demands of intelligent warehousing scenario,an improved Faster-Rcnn storage box detection model is presented based on Transformer.Firstly,the convolutional neural network Resnet50 is replaced with Swin Transformer in the Faster-Rcnn to improve feature extraction ability because the latter can extract global features and solve the problems of unsatisfactory feature extraction,redundant detection boxes and error detection boxes caused by using traditional algorithms.Secondly,a feature pyramid network is introduced to make the models suitable for multi-scale object detection.Finally,ROI Pooling is replaced with ROI Align to eliminate the errors caused by rounding the floating-point number.The model is trained on the self-built warehouse dataset.We perform random rotation,random clipping,image standardization operations for the purpose of dataset enhancements.Experimental results show that,with our proposed model,the mean average precision rate of box detection is 90.6%,the average recall rate is 93.3%,and the average detection speed is 8.9 fps.It realizes the accurate detection of storage objects and meets the needs of intelligent storage.In terms of mean average precision rate,the improvement is 6.1%,5%,10.2% and 9.7% over Fast R-CNN,YOLOv3,SSD and FCOS algorithms respectively.The recall rate is 5.9%,4%,10.1% and 9.4% higher than that of Fast R-CNN,YOLOv3,SSD and FCOS algorithms respectively.

关 键 词：目标检测 仓储场景 Faster RCNN模型 Swin Transformer模型 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]