Semantic Segmentation of Phased Array Ultrasound Images Based on Improved U-Net3+

Ultrasonic phased array images have been widely used in the welding defect detection of polyethylene gas pipelines. With the rapid development of machine vision technology, using machine assisted or automated analysis of ultrasonic images can greatly improve the defect detection speed and reduce the occurrence of human judgment errors. In the welding defect detection technology based on ultrasound images, the accuracy of image semantic segmentation is crucial for determining the defect categories and severity levels. This paper proposes an improved model incorporating residual modules and attention mechanism based on the U-Net3+ network, and applies it to the semantic segmentation of phased array ultrasonic image for defect detection in electrofusion welding. Firstly, the improved model enhances the image feature extraction ability of the encoder by adopting a residual structure between each layer of the encoder. Secondly, by introducing Convolutional Block Attention Module (CBAM) in the skip connection, the model’s utilization of original image information is strengthened, making it easier for the model to focus on effective regions in the original image. The experimental results show that the improved model has good segmentation performance on ultrasonic images of electrofusion welding, with improvements of 8.81% and 12.84% in Dice and mIoU indicators compared to U-Net, respectively. Compared to U-Net3+, the segmentation performance is improved by 1.09% and 1.81%, respectively.