Research And Application Of Rebar Counting Algorithm In Complex Scene

Deep learning has made great progress in the field of computer vision in recent years.Especially in the research and application of general object detection,great breakthroughs have been made.But in some special situations,the object detection algorithm shows some limitations.When the objects to be inspected are relatively dense,there will be occlusion between the objects.For example,rebar counting is a typical scene.Generally at the construction site,the staff in charge of acceptance shall count the steel bars on the vehicle,and the loading vehicle can only enter the site and start unloading after confirming the number of steel bars.At present,manual counting is used in the field.This method is tedious,slow and labor consuming.In view of the above situation,we can first collect pictures through the equipment,then complete the detection and counting through the algorithm of high precision,high efficiency and high robustness and modify a small amount of false cases manually,to achieve this task intelligently and efficiently.Considering the above factors,this paper proposes a new research idea,which applies the deep learning model to the identification of steel bar quantity.The main contents and innovations of this paper are as follows:(1)In this paper,a dense multi-target recognition algorithm based on full convolution network is proposed.The model uses multi-scale image pyramid as input,generates Heatmap through convolution neural network,and then fuses the multi-scale output through post-processing.The model avoids a lot of calculation of sliding window by using the characteristics of fully convolution network,and has strong interpretability.The smaller parameter scale makes it more suitable for embedded and mobile platforms,and also greatly reduces the training time,which is suitable for incremental learning.This algorithm is much better than the traditional algorithm,and its accuracy and recall rate are over 95% in the private dataset,which is 10% higher than the traditional algorithm.(2)In this paper,we propose a semantic segmentation and key point detection algorithm based on weak supervision,which can complete the detection without the anchor and NMS.The algorithm trains the semantic segmentation network through the weak supervision label,predicts the intersection and union of the objects,and performs pixel level operation on the detection results,so as to achieve the result of approximate instance segmentation.Then,the key point detection is used to distinguish the segmentation results,and the final bounding box is generated by the postprocessing algorithm.The model surpassed Retina Net,Faster RCNN and other models in datafountain competition dataset,reaching 98.8% of F1-score.(3)Using the above model,a complete intelligent counting system of rebar is designed and implemented.The user uploads the pictures firstly,after the system enhances and preprocesses the input pictures,the system detects rebar and counts quantity through the above-mentioned deep learning model,and visualizes the detection results.Users can further modify and improve the detection results through convenient operation to obtain the final count result.The two algorithms proposed in this paper can complete the detection of steel target well and have strong robustness.However,there is a problem of irregular shooting angle in the actual scene,which will lead to serious occlusion.The follow-up research can adopt the method of multi angle shooting and then splicing into a large picture,or use the laser stereo camera for data acquisition and so on.