Intelligent Recognition Mechanism Of Weld Surface Defect Image Based On Machine Vision

Today,with the rapid development of China’s intelligent manufacturing industry,welding quality is the key factor to ensure product quality.However,factors such as welding environment,welding materials,welding processes,welding power parameters,and welding operation methods may lead to various defects on the weld surface.If they are not detected,product quality will be seriously affected.The traditional detection methods have low accuracy and poor performance,and can not meet the requirements of intelligent manufacturing.The surface defect detection method based on machine vision has the characteristics of high precision,fast processing speed and intelligent processing,which is the main development trend of weld surface defect detection.The research on intelligent recognition mechanism of weld surface defect image based on machine vision is conducive to avoiding problems such as missed inspection and wrong inspection,improving the detection accuracy,efficiency and success rate,reducing the rate of defective products,and has guiding significance for the intelligent development of production process.An improved filtering algorithm based on the comparison of neighborhood absolute difference is proposed to solve the noise problem in the weld surface image.The proposed algorithm judges noise points,edge points and non noise points by calculating the absolute difference between each pixel point and its four neighboring pixels,and comparing it with the set threshold value.The proposed algorithm also obtains the pixel value that replaces the original noise point by convolving the weight matrix and the pixel matrix of the same size with the noise point as the center to filter out the noise.The proposed algorithm can not only accurately and effectively judge and filter out image noise points,but also avoid the interference of edge points,and will not filter out the edge points as noise points,effectively protecting the texture features of the weld surface image.In order to verify the effectiveness of the proposed filtering algorithm,theα coefficient method is used to establish the evaluation index system.The A-coefficients of the four surface defect images of cracks,pores,scratches and inclusions calculated are greater than 0.7,which meets the preconditions for the test.Then the filtering test is carried out for the weld surface defect image containing different concentrations of pepper and salt noise(16%,32%,48%).Then the filtering test is carried out for the weld surface defect image containing different concentrations of pepper and salt noise(16%,32%,48%).The filtering effect is evaluated and analyzed by two performance evaluation indexes,namely peak signal to noise ratio(PSNR)and mean square error(MSE).At the same time,the invariant moment feature and texture feature of the weld surface defect image are extracted for analysis.According to the analysis,the MSE value of the proposed filtering algorithm can always be controlled in a small range(18.396～22.946)for the filtering processing of weld surface defect images with different salt and pepper noise content,and the PSNR value is also very stable(34.524 d B～35.484 d B).The image has almost no distortion,and the quality is good.In order to separate the weld surface defects from the background image,an improved threshold segmentation algorithm for small defects is proposed.The proposed algorithm fully considers the similarity between the defect target and the background pixel in the same area.During the target segmentation process,the defect target is found in a small area by dividing a small area,and then it is confirmed whether the defect target is the target to be searched.An intelligent recognition system of weld surface defect image is established to detect the defect area,weld width and reinforcement of metal weld.The image processing algorithm of "median filter-corrosion operation-Canny operator" is adopted,and then the area of weld defects is calculated by segmentation of regions of interest,camera calibration,and defect contour extraction,with the measurement error less than 5%.According to the principle of laser triangulation,through structured light calibration,image preprocessing,centerline extraction based on Hessian matrix algorithm,centerline repair,sub-pixel corner detection,the weld width and reinforcement are obtained,with the maximum deviation of 0.107 mm and 0.009 mm respectively.At the same time,two kinds of software are developed,which can realize the call of basic algorithm of weld image processing,automatic recognition of weld surface defects,judgment of defect grade and cause analysis.Through depth learning,the classification and identification of weld surface defects were carried out,and the data set of 5299 pictures including four types of pores,cracks,non fusion and normal welds was established and standardized.A convolution neural network model is built,including 2 convolution layers,2pooling layers,2 full connection layers and 1 softmax regression layer,and Re LU is used as the activation function.70% of the images in the dataset were randomly selected for network model training,and the system loss value was reduced to 0.002,with an accuracy rate of 100%.The network model model.ckpt was generated.Finally,the network model is tested many times from other 30% of the data set,and the recognition accuracy is more than 95%.In view of the characteristics of scratch,crack,sand inclusion,bubble and surface depression in the surface defects of transparent weld,the equipment is built with two lighting methods of side irradiation and oblique top irradiation for image acquisition,and a feature extraction algorithm of weld surface defects based on "binary noise reduction-Gaussian filtering-binary enhancement" and a classification method based on the aspect ratio and the pixel feature of defect center are proposed.The classification and recognition of the surface defects of transparent welds are realized,with the measurement accuracy of 0.01 mm and the relative error less than 2%.Theoretical research is made on the surface defect detection method of optical element transparent weld.Gaussian smooth curve is used for filtering,and Plessey algorithm is used for corner extraction and matching to complete image mosaic fusion,defect feature extraction and area calculation.The defect area is tested by four indicators,namely,detection accuracy,image definition,detection accuracy and discrimination accuracy.The image definition is verified by the average gradient of gray scale,and the detection accuracy is verified by the presence or absence of light.The experimental results show that the accuracy of the proposed method is more than 90%.