The Research On Weld Seam Image Detection Technology For Flaw Detection Robot

Based on the flaw detection robot project in the laboratory,a set of vision-based weld seam detection system is designed,which can extract real-time deflection angle and offset of weld seam centerline.It is of great significance that improving the automation degree of flaw detection robot,detection accuracy and tracking reliability of the weld seam.The platform of system has been built from two aspects of hardware and software.The hardware has completed the design and selection of machine vision system including camera,lense and light source.The software has constructed the realization method from image acquisition,weld seam detection,centerline extraction to serial communication.The key technology including weld seam detection and centerline extraction algorithm is studied deeply in this paper,and the research adopts the traditional image processing,BP neural network and convolution neural network.In the traditional image processing,the weld seam detection is performed according to the steps of image preprocessing,edge detection,and image post-processing.In the image post-processing,by improving the existing region growing algorithm,an algorithm to remove small noise is put forward based on region growing label.And a centerline extraction algorithm based on centroid is proposed to obtain the straight line equation of the weld seam centerline.Because the traditional image processing algorithm is sensitive to illumination change and has poor adaptability,an algorithm based on sub-region BP neural network is proposed on the basis of ordinary neural network.The algorithm takes the sub-region of the image as the input and the classification result of sub-region as output of the network.Then the original algorithm is improved by proposing sub-region BP Adaboost algorithm in which sub-region BP neural network is used as the weak classifier.The experiment shows that the classification accuracy is increased by the modified BP_Adaboost algorithm.It is very cumbersome for the sub-region neural network to select the feature samples and the weld seam recognition accuracy of entire picture is much lower than that of the sub-region.The Faster RCNN algorithm based on the convolution neural network is used to detect the weld seam.The algorithm takes the entire picture as input and the location of the weld seam as output.The test accuracy can reach 90%.As for weld seam centerline extraction for the sub-region neural network and the convolution neural network,an algorithm combining the least squares method and the Hough transform based on neighborhood is proposed to obtain the straight line equations of the three types of weld seam centerline.Finally three experiments are designed in this paper.The first is the camera distance calibration experiment to get the ratio between the image pixels and the actual distance.The second experiment compares the accuracy and average running time of three kinds of weld seam detection algorithms.The third experiment proves the stability of weld seam detection algorithm based on convolution neural network.