Research Of Recognition And Counting Technique For Rebar Based On Deep Learning

With the continuous development of infrastructure construction in China,the intelligent requirements of construction sites are also increasing.In the construction site,rebar is an essential material,and the counting problem of steel bar is important in the counting management of the steel bar.In recent years,the application of deep learning to object detection and image segmentation has become a research hotspot.Aiming at the problem of counting piles,the rebar is separated from the background by image segmentation and counted quickly and accurately,thus reducing the manpower on the site and improving the efficiency.This thesis focuses on three aspects and optimizes the algorithm: Preprocessing of rebar image,segmentation of rebar image,image recognition and counting.The main work is as follows:Firstly,to improve the quality of the rebar images and facilitate the improvement of the subsequent recognition accuracy,the captured images are preprocessed.Considering that the captured images may have different sunlight intensity,shooting angles and deformation of the end face of the rebar and other factors.A quick grayscale conversion method is used to obtain grayscale images,an improved ACE method is used for image enhancement,and a fusion filter is used to reduce noise interference.From the comparison of data,the improved algorithm has better denoising effect and improves the accuracy of segmented images.Secondly,by analyzing the morphological characteristics of rebar end face images,a deep learning UNet++ network is used for semantic segmentation of reinforcement images,and the algorithm is improved.To solve the problem of poor edge segmentation and inaccurate segmentation of adhering rebar end faces,encoder-decoder networks of different levels are stacked together,and a mesh-like skip connection is used to effectively reduce the difference between contextual semantic messages.In terms of feature extraction,dilated convolution is used to replace the original convolution to expand the receptive field of feature extraction and reduce computation.In terms of feature output,connecting the extracted multi-level features can obtain multi-level fusion features and better understand the semantic messages at different levels of text.Experimental results show that this method can achieve excellent segmentation results.Finally,by analyzing the characteristics of the connected region of the bar binary image,a classification and counting method for the rebar is designed.Based on the traditional recognition method,the counting method of multi-feature fusion is used to classify the rebar,and the incomplete end face,the single root end face,the adhesion end face and the noise are identified respectively.The recognition and counting experiments of rebar end images with different characteristics show that the multi-feature fusion classification and counting method can effectively improve the accuracy of rebar counting.