Automatical Identification Technology Research Of Weld Defect Based On Neural Network

With the rapid development of ray detection and image processing technology, weld defect detection is also gradually changed from manual operation to intelligent operation based on computer. Analizing and identifying digital weld image on computer not only improves the detection efficiency and economic benefits, but also has a key advantage of operating easily. Therefore, how to improve intelligence, automation and quantitation of ray detection is a hot issue in the current research of ray detection area.There are a lot of features for original ray image of weld detection such as narrow gray scale range, ambiguous defect edge, more noise in image, disappeared defect feature, and so on, which are influencing effects for comprehensively analizing and evaluating ray image detection. Therefore, how to extract the defect and features used to identify image and identify classification, rapidly and exactly is a major prolem in this research topic.Taking the requirement of real engineering project as background, this paper studies weld film image, in which the defect and feature of image is extracted using image processing method. An intelligent and automatical identification system of weld defect, with GUI interface, is designed by MATLAB. Different types of algorithm module are used and simulated in this research, from which the optimal algorithm is choosed to identify weld defect. The main contents of this research are introduced as the following. First, based on average gray value of weld film, the image is judged to be eligible or not. Second, adaptive weighted median filter is used in image filter to remove the noise of image. Third, fuzzy theory is used to enhance image, which not only enhances image brightness, but also makes edge of weld and defect be more obvious. Fourth, morphologic characteristics are used to extract defect, from which disturbance of noise can be removed in the fullest extent. Fifth, subtraction angiography is used to extract defect. The defect in weld film image is removed through mean filter and an ideal simulated weld is created, which will be subtracted by original image to get defect. Sixth, feature parameters are calculated based on aim feature of detected image, according to assumption formula of possible target in image. Seventh, a kind of modified learnling algorithm is used to design RBF neural network to identify defection, which is used to accomplish samples training. Finally, RBF neural network is used to identify defect successfully. The research can not only increase the efficiency of identification, but also decrease cost, which meets the necessary request economic and social development.MATLAB is used to simulate image processing algorithm and neural network in this paper. With object-oriented visual GUI interface, stability weights are trained and performance is tested through a large quantity of training samples based on weld image wih defect. Experiments have shown that the modified RBF neural network can identify weld defect types of line type and the T type under suitable feature parameters and correct calculation. From the modified mehod, the position of weld defect can be marked correctly and feature parameters of defect are displayed. The final accuracy of identification can reach92%, the result of which is displayed through GUI interface.