Research On Weld Trajectory Segmentation Method Based On Fully Convolutional Networks

Welding technology occupies a very important position in contemporary industrial manufacturing,and the identification and detection of the weld seam position is the most important prerequisite for the work of automated welding systems.Only by accurately detecting the position of the weld seam track can the machine reasonably plan the welding path,thus ensuring the normal operation of the entire welding system.In the actual welding manufacturing process,the detection of weld seam position has been affected by many factors,such as the complex environment in the factory,the type of welded workpiece material,the shape and size of the weld seam,etc.It is difficult to achieve the expected detection effect by relying on traditional image processing algorithms at this time.Therefore,how to improve the accuracy of weld track position detection and the robustness of the algorithm has become the focus of research in the field of automated welding.With the improvement of computer calculation performance in recent years,the use of deep learning technology to detect weld position has become a research hotspot in the field of welding.In this paper,deep learning technology is introduced to analyze and study the image semantic segmentation model of the full convolutional neural network,and a welding seam trajectory segmentation algorithm model based on the improved full convolutional residual neural network is proposed.Improvements and innovations are proposed from the feature extraction network and the convolutional layer structure of the original model.The first aspect is that Residual network is used in feature extraction network,which can effectively solve the problem of network gradient disappearance and gradient explosion.At the same time,residual structure in the network can accelerate the convergence rate of the network model and improve the overall performance of the model.The second aspect is that dilatation convolution is introduced into a Bottleneck in the residuals structure to replace ordinary convolution.This method,while bottleneck enlarges the convolution kernel receptive field,reduces the number of parameters in the network.The method in this paper can effectively improve the rough edge segmentation effect and small target loss in the task of welding seam trajectory segmentation,and help the model to better extract the details of the welding seam image.The above research was synthesized to design and develop a Py Qt-based human-computer interface software that can be used directly by users independently of the complex Python deep learning environment.The software visualizes the entire weld trajectory segmentation process by loading the improved model and weight files of this paper,and its results are saved in the user’s pre-defined path,thus completing the weld inspection task efficiently and conveniently.Finally,to verify the feasibility of the method,a new Weld Seam Dataset is established in this paper,and the improved model is experimentally compared and analyzed on this dataset.The MPA of the improved model is improved to 94.8% and MIo U is improved to 90.6%,and the experimental results prove that the method in this paper has higher performance and can achieve the expected results.