Research On Key Technology Of Rebar Surface Defect Detection Based On Deep Learning

As an indispensable building material in civil engineering,the production demand of threaded steel in China is increasing day by day.During the rolling process,it will be affected by many factors,such as material problems,damage to rolling equipment,surface impurities and so on,resulting in a variety of defects,which seriously affects the economic benefits of iron and steel enterprises.At present,some enterprises still use manual visual inspection and traditional image processing methods to identify and detect defects on the surface of threaded steel.However,due to subjective errors in manual visual inspection,and traditional image processing methods need to rely on specific tasks and artificial set characteristics,the recognition rate is low and the detection effect is not ideal.There are also many limitations,so this subject proposes a threaded steel detection scheme based on deep learning.By applying this product,the production enterprises can reduce labor output,reduce production costs and improve economic benefits,thus promoting the development of steel production industry.The main work of thesis is as follows:1.Firstly,this thesis makes a brief analysis on the types and causes of surface defects of threaded steel,lists the current mainstream algorithms and frameworks for defect detection and identification at home and abroad,and selects Faster RCNN as the basic frame for surface defect detection of threaded steel after comparison.In order to solve the problems of uneven size and distribution of surface defects on threaded steel,the anchor area generated by RPN and the selection ratio of Io U are improved to better adapt to the characteristics of surface defects on threaded steel studied in the thesis.2.Due to the large number of training samples required for in-depth learning and the small number of data samples for surface defects of threaded steel obtained in this paper,the method of transfer learning is selected to solve the problem of small data samples,and the defect detection model is pre-trained and fine-tuned with a relatively large number of steel databases.The pre-training weight is saved as the initial weight for training the threaded steel data set,and then the labeled and divided surface defect data set of threaded steel is input into the detection model for training.After adjusting the learning rate and setting the parameters,the model is trained.The experimental results show that the detection algorithm in the thesis meets the requirements of less iterations and higher detection efficiency.3.Although the migration learning improves the detection efficiency of the detection system,there are still some differences between the pre-trained data set and the surface defects of threaded steel.In order to further improve the accuracy of the defect detection system and solve the problem of uneven sample distribution in the threaded steel data set,the method of data enhancement with small samples is analyzed in this paper.Considering that the data of surface defect of threaded steel obtained during the experiment is too little and some defect characteristics are not representative,thesis introduces two widely used data expansion algorithms,and respectively inputs the data generated by supervised data enhancement and unsupervised data enhancement into the network for training and compares the accuracy of training results.Experiments show that unsupervised data enhancement method is more applicable to the data set in the thesis,because the unsupervised data enhancement method is more suitable for the training of deep learning network because the shape,size,position and size of defects generated by it are more random than traditional image processing algorithm.Finally,the DCGAN algorithm of countermeasure learning is generated by deep convolution in unsupervised data enhancement to balance data set and expand defect samples of threaded steel.4.During the training,it is found that defects such as scars or small bubbles account for a small proportion in the whole picture,which leads to unsatisfactory detection effect of some defects in the detection process.In this thesis,the training method of FPN + Faster RCNN is used to increase the multi-scale semantics information of features to improve the accuracy of the algorithm.The experimental results show that the detection model with pooling pyramid(FPN)can greatly improve the recognition accuracy of small targets.