Research on Weld Defect Recognition Method Based on Mask R-CNN

Abstract

Automatic detection of weld defects is of great significance to ensuring the quality of industrial products and engineering safety. To address the challenges of low contrast, large defect scale variation, complex background noise, and limited sample data in ultrasonic weld defect images, an improved Mask R-CNN instance segmentation model is proposed. First, the original ResNet+FPN backbone network is replaced with the RSU7 multi-scale feature extraction module to enhance the model’s ability to capture details of tiny defects through a nested U-structure and residual connections. Second, the CBAM attention mechanism is connected in series between the backbone network and the region proposal network to suppress background noise and highlight defect regions in both channel and spatial dimensions. Experiments are conducted on an ultrasonic weld defect dataset containing only 105 images. The results show that the improved model achieves a mean average precision (mAP) of 0.7564, which is 4.2% higher than that of the baseline Mask R-CNN (ResNet50+FPN). The coefficient of variation (CV) is 2.26%, indicating better training stability than the comparison models. The maximum AP drop rate under image disturbance is only 9.1%, demonstrating strong robustness. The proposed method realizes high-precision and high-stability defect detection and segmentation in small-sample scenarios, providing an effective technical solution for industrial ultrasonic welding quality inspection.