Research On Reinforcement Counting Method Based On Lightweight Neural Network

Reinforcement is one of the basic functional materials in steel enterprises,which is widely used in the construction industry.Research on the automatic statistics method of steel bar quantity can speed up the pace of realizing the intellectualization of iron and steel enterprises.In the past two years,in computer vision,deep learning has made great progress.The counting method based on object detection has been gradually applied in industrial detection.However,the requirements for hardware resources are high,the network parameters are large,and the forward inference speed is slow,which cannot meet the requirements for real-time steel bar detection and counting in embedded devices with low computing power.Therefore,on the basis of previous research,this thesis designs a lightweight network model structure with both speed and performance from the perspective of lightweight feature extraction and feature fusion,which can accurately and efficiently detect and count the steel targets.The main work is as follows:(1)Aiming at the problem of insufficient data set of steel bar end face image,the required data set is made,and the algorithm comparison experiment is carried out on the self built data set using common evaluation indicators.Based on the experimental results,YOLOv4 algorithm,which gives consideration to both the detection speed and accuracy and has good open source performance,is selected as the basic algorithm to realize the reinforcement counting task.(2)The number of network layers in YOLOv4 is deep,and a large number of ordinary convolutions are used,resulting in a large model volume and consuming a large amount of computing resources.To solve this problem,the backbone of the network model uses a lightweight neural network Mobile Netv2 to extract image features.The rest of the model structure remains the same as YOLOv4.The basic unit module uses depth to separate convolutions,so that the model becomes lightweight,The experimental results show that the lightweight network model has the characteristics of low complexity and small volume.However,compared with YOLOv4,the precision of the built lightweight network has declined,so further improvement is needed to improve the detection precision.(3)In order to enhance the ability to perceive the spatial position information of dense steel targets,coordinate attention is introduced into the reciprocal residual module of Mobile Netv2 for targeted optimization;In order to make better use of the semantic information of shallow,middle and deep feature maps and enhance the recognition ability of the model for multi-scale objects,a lightweight feature fusion network LFPN is designed to improve the detection efficiency and accuracy;In view of the particularity of the foreground and background in the reinforcement dataset image,Focal Loss loss function is introduced into the confidence loss to balance the positive and negative sample proportion of the reinforcement dataset.In view of the dense characteristics of the bundled reinforcement targets,EIo ULoss is introduced as the boundary box regression loss function to solve the problem that the boundary box between the bonded reinforcement is not easy to identify and enhance the regression ability of the boundary box.The experimental results show that compared with Mobile Net-YOLOv4,the AP value of the improved network is increased by 8.41 percentage points,and the number of parameters is reduced by 51.3%;Compared with YOLOv4,the AP value has increased by 2.58 percentage points,the amount of parameters and computation has decreased significantly,and the FPS has increased by3 frames/second.The automatic counting method of bundled steel bars studied in this thesis can be applied to counting in the process of steel bar production and transportation,and has certain practical application value.