Abstract: To accurately and quickly extract the molten pool image in the welding process,a molten pool image segmentation method based on a lightweight DeepLabV3+ network was proposed.The backbone network of DeepLabV3+ was replaced by the optimized MobileNetV2 network from Xception to reduce the number of model parameters.The Coordinate Attention (CA) mechanism was introduced to improve the model's ability to extract the molten pool image.The training method of transfer learning was used to solve the scarcity of molten pool samples and improve the accuracy and generalization ability of the model.The experimental results showed that the Mean Intersection over Union (MIoU)of the improved model under the molten pool data set was 94.65%,the Mean Pixel Accuracy (MPA) was 96.67%,the inference time of a single picture was 11.09 ms,and the model parameter quantity was 5.81 M.Compared with classical networks such as SegNet,PSPNet,UNet and DeepLabV3+,the improved algorithm had a smaller number of model parameters,shorter single-image inference time and maintains a higher mean intersection over union,which could better balance the image segmentation accuracy and real-time performance.

Key words: semantic segmentation, DeepLabV3+, lightweight, molten pool

摘要： 为了准确快速地提取焊接过程中的熔池图像,提出一种轻量级DeepLabV3+网络的焊接熔池图像分割方法。首先,将DeepLabV3+的主干网络由Xception替换为优化后的MobileNetV2网络以减少模型参数量。其次,引入坐标注意力(CA)机制,提高模型对熔池图像的提取能力。最后,利用迁移学习的训练方法,解决熔池样本稀缺的问题,并提升模型的精度和泛化能力。实验结果表明,改进后的模型在熔池数据集下平均交并比(MIoU)为94.65%,平均像素精度(MPA)为96.67%,单张图片推理时间为11.09ms,模型参数量为5.81M。与SegNet、PSPNet、UNet和DeepLabV3+等经典网络相比,改进后算法的模型参数量小,单图推理时间较短,且保持较高的平均交并比,能够更好地平衡图像分割精度和实时性。

关键词: 语义分割, DeepLabV3+, 轻量级, 熔池