Rebar Localization And Segmentation Based On Convolution Neural Network

With the advent of the era of deep learning,counting methods based on deep learning are gradually applied.Current visual methods face many difficulties and challenges when processing reinforcement images for counting.At the construction site,workers manually count the number of rebars,which is laborious and time-consuming(sometimes several hours).Since the images captured from the construction site vary according to site conditions,there are some problems,such as irregular End shape,uneven illuminance,uneven color and overlapping end faces,etc.All these factors lead to unstable recognition results when using current image processing algorithms.Traditional machine vision counting strategies do not perform well in terms of speed and accuracy.Therefore,it is of great significance and great practical application prospect to carry out the research on the topic of the positioning and segmentation of steel bar end faces.Based on the convolution neural network as the research method,this paper focuses on the positioning and segmentation of the end face of the steel bar.Through reasonable and optimized model design,efficient positioning and segmentation of the end face of the steel bar is achieved.The research results are summarized as follows:1.A data enhancement method based on sliding window(SWDA)is proposedIn order to solve the problem of insufficient data in the training process of the current reinforcement learning algorithm based on deep learning,this paper proposes a sliding window-based data enhancement method(SWDA).This method includes the process of data reading,the generation of the white mask area,the generation of the available area based on the sliding window,and the backfill process of the target pixels and labels.The proposed data enhancement method for sliding windows has two advantages: first,it is suitable for filling objects of any size,which is helpful to increase the number of objects to be detected in the image;second,the method uses a random selection method,which is beneficial to Increase the diversified information of the target and improve the generalization ability of the detection and positioning model.2.Proposed a Fibonacci incremental mask labeling method(FIMLM)In order to alleviate the problem that it takes a lot of manpower and time to produce image segmentation data sets in deep learning,it takes 30 minutes to create label data for a 2666 × 2000 image using PS.Therefore,this paper proposes a Fibonacci Incremental Mask Labeling Method(FIMLM)for labeling the segmentation data set of the steel bar end face.This method belongs to a semi-automatic labeling method.In the labeling process,manual masking for labeling errors is introduced to improve the quality of model labeling.After the optimization of the iterative training of the model,the coincidence degree of the prediction mask of the model is getting higher and higher,and the quality of the annotation is also improved.3.Proposed a steel bar detection and positioning model(Inception-RFB-FPN)Aiming at the problems of accuracy and positioning speed in the steel bar positioning algorithm,this paper proposes a steel bar detection and positioning model(Inception-RFB-FPN).The model contains an information retention layer,data abstraction layer(Inception),RFB-FPN image feature pyramid module and detection layer.The advantage of the proposed steel detection and positioning model is that it takes into account the positioning accuracy while maintaining the real-time detection speed(single image detection time is 0.0306s),which is very suitable for mobile application scenarios.4.The performance of three types of fully convolution networks(FCN)for the segmentation of the end face of the steel bar is studiedThe mask prediction of the high-resolution image segmentation model will gather in the neighboring target,making the mask boundary of the target unclear.This paper decomposes the target into small images,uses single target segmentation and finally merges and maps back to the original image,which alleviates the problem of mask condensation to a certain extent.The input image size of FCN uses the approximate size 128 of the statistical average.The inference time of 128 × 128 pixels for a single VGG16-FCN is 2.6156 ms,3.2635 ms for ResNet18-FCN,5.3670 ms for ResNet34-FCN.