Research On Feature Extraction And Defect Recognition Of Weld Surface Image

The development of modern engineering requires speed,reliability,and low cost in the detection of the surface-welding technology.This paper mainly studies the artificial intelligence defect-detection technology of surface-welding based on image processing,to realize the accurate detection and classification of defects in surface weld by edge extraction,weld location,feature extraction,neural network classification and other methods.Key technologies are as follows:First,the Canny edge extraction algorithm is improved to realize identification and localization of the weld joints.Due to the problems of Canny algorithm,k-nearest neighboring filter mean is used to replace Gauss filter,and a new gradient template is constructed based on Scharr operator to calculate image gradient;after the suppression of non-maximum value,Otsu algorithm with inhibition factor is used for gradient map,which makes adaptive threshold more reasonable and edge positioning accuracy improved.The experimental results show that the algorithm can be effectively applied to weld identification and location.Second,combining the texture and shape features of weld image,a new feature vector of weld surface quality is established,and then the BP network identification and classification of welding surface defects is achieved.Comparisons of the image features were made,of the welding beads,burning-through,incomplete welding,surface porosity and,the normal welding.In order to reduce the impact of noise,five texture features of weld surface image,i.e.contrast,correlation,energy,entropy and homogeneity,and four shape features such as area,perimeter,circularity and minimum external matrix aspect ratio,were extracted to construct welding surface image feature vector The strong learning ability of P neural network realizes the classification of welding surface defects.The experimental results show that the method can effectively identify the welding surface defects,and the accuracy of the test sample reaches 93.4%.Thirdly,in view of the difficulty of image feature vector extraction due to complex image noise interference and large sample size,the advantages of convolution neural network adaptive feature extraction and big data sample processing ability are introduced.The AleNnet convolution model is used to process the image data set of welding surface defects after image enhancement to extract the deep abstract features of the image.In order to verify the identification results,the classification results of performance index evaluation are calculated from the two aspects of accuracy and loss rate.Conclusion is that the recognition rate of convolution neural network is higher,which is stable at 95.2%±1%.Finally,the software system of feature extraction and defect recognition of weld surface image is designed,and two kinds of identification methods are embedded for different situations.The system mainly includes three modules:image input,positioning and defect identification,and result output.The establishment of the platform realizes the visualization of weld inspection process as well as image acquisition,import and storage.The debugging results show that the designed software system is simple,effective and operability.