\n
\n
\n
\n
\n
\n
\n
\n
\n
\n 面向焊缝X射线图像缺陷的两阶段检测方法
Two-stage detection method for defects in weld X-ray images
\n
\n
\n
\n
\n -
\n
\n
-
\n \n \n\n\n 摘要: \n 为了解决传统X射线焊缝检测依赖人工判读、效率低且精度不足的问题,文中提出一种基于深度学习的端到端两阶段检测方法.阶段1采用YOLOv11模型提取焊缝区域掩膜,通过仿射变换与滑动重叠裁切消除方向差异并减少背景干扰.阶段2采用Rd_evit模型对裁剪后的子图像实现缺陷像素级分割,模型中设计了Dyf模块增强多尺度特征提取能力,通过Rsc模块优化跳跃连接,同时,构建Dice系数损失、边界损失和聚焦损失的组合损失函数,以解决小目标、边界模糊和类别不平衡问题.最后,通过子图像融合策略处理重叠区域的预测冲突,并经逆仿射变换将分割结果映射回原始图像坐标系,实现端到端的缺陷检测.阶段1采用\n \n
1640 张焊缝区域标注图像,YOLOv11的Dice系数达到85.12%;阶段2试验采用含裂纹、未焊透、未熔合、夹铜/夹钨和圆形缺陷5类缺陷的2719 张X射线焊缝图像,结果表明,Rd_evit的Dice系数达69.87%、交并比指标达58.55%,较基线模型EViT-UNET分别提升11.67%和15.53%,其中未熔合类缺陷检测精度提升27.15%. \n
\n \n\n \n\n\n Abstract: \n To address the issues of low efficiency and insufficient accuracy of traditional X-ray weld seam detection which relies on manual interpretation, an end-to-end two-stage detection method based on deep learning was proposed. In the first stage, the YOLOv11 model was employed to extract the weld seam region mask. Affine transformation and sliding overlapping cropping were then applied to eliminate orientation differences and reduce background interference. In the second stage, the Rd_evit model was used to achieve pixel-level defect segmentation on the cropped sub-images. Within this model, a Dyf module was designed to enhance multi-scale feature extraction capability, and an Rsc module was utilized to optimize skip connections; additionally, a combined loss function integrating Dice coefficient loss, Boundary loss, and Focal loss was constructed to address the issues of small targets, blurred boundaries, and class imbalance. Finally, a sub-image fusion strategy was adopted to resolve prediction conflicts in overlapping regions, and the segmentation results were mapped back to the original image coordinate system through inverse affine transformation, enabling end-to-end defect detection. In the first stage, 1 640 annotated images of the weld seam region were used, and the Dice coefficient of YOLOv11 reaches 85.12%. In the second stage experiment, 2 719 X-ray weld images containing 5 types of defects, namely cracks, incomplete penetration, lack of fusion, copper/tungsten inclusions, and circular defects, were used. The results show that the Dice coefficient of Rd_evit reaches 69.87% and the intersection over union indicator reaches 58.55%. They are improved by 11.67% and 15.53% respectively compared with the baseline EViT-UNET model, and the detection accuracy of lack of fusion defects is improved by 27.15%. \n\n \n
\n \n
\n
- \n \n \n \n \n \n
- \n \n \n \n \n
- \n \n \n \n
- \n \n \n \n
\n
\n 摘要
\n
\n
\n
\n
\n HTML全文
\n
\n
\n
\n
\n 参考文献(27)
\n
\n
\n
\n
\n
\n 相关文章
\n
\n
\n
\n
\n
\n
\n
\n
\n 施引文献
\n
\n
\n \n
\n
\n
\n
\n
\n
\n 资源附件(0)
\n
\n
\n
\n
下载:
\n\t\t\t\t