Rebar Counting Algorithm Research Based On Convolutional Neural Network

Rebar is one of the most basic materials used in various types of construction.Research of rebar counting algorithm is of great significance to improve the automation level of rebar industry.This thesis researches the counting problem of bundles of rebars on construction sites and the automatic packaging counting of rebars on the production line.Based on the convolutional neural network,the static rebar counting algorithm for the picture of bundled rebars' end face,and the dynamic counting algorithm for the infrared hot rolled rebars' end face video on the production line is proposed,and the infrared hot rolled rebar video intelligent counting system is designed and implemented.The main research work of this thesis can be summarized as follows:First,a static rebar counting algorithm based on a cascaded region proposal head network is proposed.Due to the low proportion of each rebar end face in the picture of the rebar,a rebar end face object detection algorithm based on cascaded region proposal head network is proposed.First,feature pyramid fusion algorithm is used to extract multi-layer feature maps from feature extraction network,and then extracts region proposals.Finally,The region proposals are progressively predicted and regressed by using the three-level cascaded region proposal head network,and the result is post-processed using soft NMS to obtain the end face object detection result.In the training,the balanced hard negatives sampling and the balanced L1 loss function are used to further improve the detection accuracy of the network.Second,an infrared hot rolled rebar video counting algorithm based on lightweight single-scale feature map network is proposed.Based on the requirements for dynamic counting of hot rolled bars on conveyor belts in the field of view of a fixed infrared camera,the algorithm sets a boundary line at a fixed position in the video,detects the rebars and tracks them,and counts the rebars passing through the dividing line.In terms of detection,a lightweight single-scale feature map detection network is designed.A depthwise separable convolution network with small computational complexity is used as the feature extraction network.A single-scale feature map is designed according to the feature of the end face of the rebar.Anchors are set on the feature map to detect the rebar end face.In terms of tracking,this thesis uses Kernel Correlation Filter(KCF)algorithm to track the detected rebar objects in video frames.The experiment results show that the proposed dynamic rebar counting algorithm not only has good real-time performance,but also has counting accuracy,which fully meets the requirements of practical applications.Third,the infrared hot-rolled rebar video intelligent counting system is designed and implemented.Based on the Qt IDE and the tensorflow deep learning framework,the lightweight single-scale feature map rebar detection network is deployed to the Windows platform.Combined with the KCF object tracking algorithm,dynamic counting of the input infrared hot rolled rebar video is realized.Using enlarged tracking box and the inflow rebar object discrimination mechanism,to eliminates the counting error caused by the height similarity of the rebar end face object and the belt vibration.