Research On Package Detection Algorithm Based On Convolutional Neural Network

With the rapid development of e-commerce,logistics demand is higher and higher.Intelligent logistics system plays an important role in improving the operation efficiency of logistics transit center and reducing the management cost,so it is widely concerned.Package detection is one of the key technologies of package management in intelligent logistics system,and it is also one of the specific applications of target detection in the field of logistics.Although the target detection technology has made great progress in recent years,it will face great challenges to directly apply the existing target detection technology to the package detection task.On the one hand,a large number of packages result in a dense package imaging target;on the other hand,because of the long imaging distance,the package imaging shape is small.In addition,various transmission related devices around the package will cause complex background interference.These challenges make package detection a difficult task.For this reason,this thesis first reviews the research status of object detection;secondly,based on the monitoring video data of Wuxi Distribution Center of China Post Group Corporation,this paper makes an in-depth study on the package detection algorithm based on convolutional neural network,and the main work is as follows:(1)According to the data requirements of package inspection,the package inspection data set is constructed.Through the monitoring video of the logistics transfer center provided by Wuxi Distribution Center of China Post Group Corporation,the image is preprocessed,and the packages in the image are manually marked with labelimg opensource software to find out the area of the packages in the image and draw the border box.Thus,a package data set representing the actual situation is constructed.(2)Aiming at the problem of dense packages in Logistics Transfer Center,an improved fast-rcnn package detection algorithm is proposed.Based on the classic target detection model faster r-cnn,the regression term in the package detection algorithm is adjusted by using the repgt loss function.In the case of dense packages,the prediction frame deviates from the non current target frame.Based on the original target detection algorithm fast r-cnn,the detection accuracy of the algorithm is improved up to 2.38 ap,and the parameters in the loss function are analyzed,and the influence of parameters on the overall algorithm is discussed.(3)In order to solve the problem that the information of small target wrapped in convolution operation is lack and it is difficult to detect,based on the previous research,multi-scale feature fusion method is used to fuse the shallow feature map of convolution neural network.Through the experiment,it is found that choosing the appropriate method of anchor and feature fusion can effectively improve the detection accuracy of small target package detection.Finally,the detection accuracy(AP)is increased from 59.98 to 68.74,which improves the detection accuracy of small target packages.