Semantic segmentation algorithms based on deep learning can segment an image of the molten pool generated by the laser welding process. The results of an analysis of this molten pool can be used to adjust the welding parameters in real time, thereby improving the quality of laser welding. Although it is desirable to use a simple neural network to extract the complete molten pool, the existence of the wire, arc, spatter, smoke, and other sources of interferences often make the welding environment very complex. Thus, it is very difficult to extract the complete molten pool and detect it using a simple neural network. This study addresses the problems with the welding process. The use of denoising and tracking networks are proposed for a molten pool enhancement method, along with the use of a semantic segmentation network based on a Mask2Former lightweight improvement to accurately segment a molten pool image.

This study analyzes the noise in a molten pool image from a dataset. The causes and types of noise in molten pool images vary and a clear image of the molten pool may be lacking. Therefore, SCUNet is adopted to realize the blind denoising of a molten pool image. The molten pool is the smallest target in the image. Therefore, in order to reduce the interference from the useless parts of the image and improve the detection efficiency, we utilize MixFormer to track and localize the molten pool during the welding process. In order to meet the demand for the real-time detection of the molten pool, the Mask2Former semantic segmentation network is lightened and improved, and MobileNetV3 is used instead of the original backbone network. Then, ResNet101 is used as the backbone of the Mask2Former network as a teacher network to distill the knowledge of the lightened network, allowing the accurate segmentation of the molten pool image.

An evaluation of the denoising network index and the denoising results show that the algorithm reported in this paper can significantly reduce the interference caused by the arc light during the laser welding process compared with other algorithms. Thus, the changes in the molten pool image are smoothed, and the molten pool pattern is clear (Table 1 and Fig. 7). A comparison shows that MixFormer is better able to balance the tracking accuracy and speed compared to other networks (Table 2 and Fig. 8). A performance evaluation of the backbone network using a test set shows that the detection accuracy of the MobileNetV3 model is slightly lower than that of the other selected models, but its number of parameters and computation amount are 0.93×10⁶ and 0.32×10⁹, respectively, which are smaller than the numbers for the other selected networks. Its M_IOU can reach 97.02%, which shows that MobileNetV3 can ensure the feature extraction ability of the network while keeping the number of parameters and computation amount small. This indicates that MobileNetV3 ensures the feature extraction ability of the network while keeping the number of parameters and computation volume small (Table 3). The results of an experimental analysis of the effect of temperature on distillation show that when the temperature is set to five, the IOU and M_IOU reach their maximum values, indicating that the network has the best distillation effect (Table 4). The inference time of the method proposed in this paper is 84.7 ms, and the M_IOU is 97.21%, indicating that the distilled model can better balance the accuracy and real-time performance, and has good detection performance (Table 6 and Fig. 14)

This paper reports how the molten pool generated by the laser welding process was used as an experimental object. Using the reported method, the image of the laser welding molten pool is blindly denoised by SCUNet, and then the molten pool is tracked using the single-target tracking algorithm. Finally, the molten pool is segmented using Mask2Former. The weight of the segmentation network is lightened using the knowledge-distillation method. Thus, MobileNetV3 can dynamically segment a molten pool through knowledge distillation by learning the knowledge of the teacher network. The experimental results show that (1) the peak signal-to-noise ratio and structural similarity of 37.87 dB and 0.95, respectively, are better than those of a traditional denoising algorithm when SCUNet is used for denoising the molten pool image, indicating that this algorithm can realize the denoising of an image under the guarantee of image similarity. (2) The MixFormer algorithm is used to track the molten pool. Compared with STARK and SiamRPN++, it achieves better results in realizing molten pool tracking. (3) Compared with the classical semantic segmentation model, the M_IOU and detection speed of the lightweight molten pool detection method based on the improved Mask2Former are 94.21% and 84.7 ms, respectively, which allow it to quickly and accurately detect the molten pool. The smaller number of parameters for the backbone network is more favorable for real-time detection and deployment, which shows the superiority of the method reported in this paper in the detection of the molten pool produced by laser welding.