Pre-Welding Defect Detection in Hull Construction Based on Dual-Branch Attention Feature Enhancement

In hull construction, manually identifying and assessing pre-welding defects is a challenging and time-consuming task, with potential defects posing significant safety risks. Fortunately, with the advancement of deep learning, convolutional neural network (CNN)-based recognition models have been successfully applied to computer vision tasks. However, due to the diversity of interfering factors in pre-welding defect images, current models struggle to capture critical defect features. When using attention mechanisms to address this issue, attention shifts can occur, leading to lower model recognition accuracy. To address the issue of attention shift, this paper proposes a Dual-Branch Attention Enhancement module (DBAE). This module extends traditional attention mechanisms by introducing an additional branch focused on enhancing defect features, guiding the attention towards these critical features, and improving feature extraction. Experimental results demonstrate that the DBAE module has superior feature capture capabilities compared to other classic attention modules, with the most significant improvement reaching 7.25 %, consistently outperforming other attention mechanisms. Furthermore, the results confirm that DBAE is effective across various classic classification networks, exhibiting good generalizability.