Research On Defect Recognition Of Weld Image Based On Deep Learning

As a basic connection method between workpieces,the quality of welding directly affects the performance of the structure.With the rapid development of industry and the increase of competitive pressure,more and more attention is paid to the welding quality.Therefore,professional interpretation and evaluation of X-ray images of welding seams become extremely important.Because of the influence of many subj ective factors,manual film evaluation is gradually eliminated by people.In addition,the advent of the big data era makes the acquisition of data becomes easier,which slao makes the work quota of manual film evaluation becomes greater.Therefore,the study on automatic identification of defects in X-ray images of welds has received extensive attention.At present,the automatic defect recognition technology mainly relies on three steps:traditional image processing,manual feature extraction and pattern recognition.However,the manual feature design relies on professional knowledge and experience,and the operation is tedious,and there are many components of manual intervention,so it is difficult to achieve intelligence actually.In addition,there may be a large number of features,thus the feature selection is usually required.Finally,the features are used to be classified by traditional pattern recognition method.These major steps are often performed independently and are difficult to optimize jointly.In recent years,with the rapid development of computer vision technologies,such as sparse representation and deep learning,are opening new avenues in the research of automatic object recognition from optical images.In particular,the great success of deep learning in natural image recognition and classification prompts people to try to apply it to industrial detection.However,at present,the application of deep learning in X-ray images of welding seams is relatively rare.This paper takes X-ray images of welding seams as the research object,further studies the algorithms of sparse representation and deep representation,probes into the feasibility of their recognition and classification on X-ray images,and proposes two end-to-end automatic recognition models based on deep network for recognizing and classificating the weld defects.This paper first describes the research background and significance of weld defect detection,introduces the image recognition technology and the basic theory of machine learning,then elaborates the theory of the back propagation algorithm in the traditional neural network,and puts forward the difficulties when appling the algorithm in the deep network.Then it analyzes the opportunity of the rise of deep learning and the three waves of its development.This paper also introduces the theory of sparse representation,including solution method of sparse representation,the algorithm of dictionary learning and the classification based on sparse representation.The classification method based on sparse representation solution was applied to the classification of welding defect images and achieved high accuracy,which also proved the advantages of sparse representation.Then,the theories of the principal component analysis algorithm and autoencoder are introduced.The principle of learning low-dimensional features is described,and the ability of extracting features and reconstructing images from handwritten digital images is studied.In this paper,based on the good expression ability of sparsity representation,the theory of the sparsity autoencoder and its ability to extract features are introduced in detail.The features extracted using different optimization algorithms are discussed.Then,on the basis of single-layer representation learning,a deep autoencoder network is established.The influence of the structure and parameters of network on the classification accuracy is studied.The defect recognition of the whole image is realized by combining the deep network and sliding window method.Finally,this paper introduces three solutions to the problem of unbalanced classification.In this paper,an end-to-end deep convolutional network is proposed to classify and identify defects.In addition,this article also explores the feasibility of appling the transfer learning to the defect classification in weld image.The results show that the deep convolution network has the strong ability in learning image features,thus it can be applied to the problem of this article.The proposed network has a higher classification accuracy and computational efficiency,thus it has great practical value.