高级检索

多级多尺神经网络自搜索的焊缝缺陷语义分割

Semantic Segmentation Method of Weld Defects Using Multilevel and Multi-scale Neural Network Self-search

  • 摘要: 为进一步提高焊缝缺陷X-ray底片低质影像语义分割精度, 降低人工设计网络主观影响及耗时等问题, 提出一种多级多尺神经网络自搜索的焊缝缺陷语义分割方法. 通过对多尺度轻量化候选操作、通道注意力机制和多层级动态神经网络架构的设计, 从不同维度提升网络对低质影像缺陷特征提取的表达能力; 同时对网络训练早期与最终识别性能之间的潜在关联探索, 提出使用固定采样逐步确定最优候选操作的渐进式快速神经架构搜索方法. 在架构搜索阶段使用自行采集、标注的483张X-ray焊缝缺陷图片经过随机裁剪、旋转、平移等数据增强操作进行架构寻优, 最终以较低的搜索成本自动构建出焊缝缺陷语义分割网络. 实验表明, 所提方法对X-ray焊缝缺陷进行语义分割最终mIoU指数达到了49.23%, 高于人工设计网络的45.41%和直接使用模型迁移的28.86%, 网络自搜索速度和分割效果提升明显.

     

    Abstract: To further improve the accuracy of semantic segmentation of low-quality X-ray images of weld defects and reduce the subjective impact and time-consuming of artificial design networks, a semantic segmentation method of weld defects using multilevel and multi-scale neural network self-search is proposed. Through the design of multi-scale lightweight candidate operations, channel attention mechanism and multi-level dynamic network, the expression ability of the network to extract defect features of low-quality images is improved from different dimensions; At the same time, through the exploration of the correlation between the recognition performance of the early and final stages of network training, a gradually fast neural architecture search method using fixed sampling to gradually determine the optimal candidate operation is proposed. In the architecture search phase, 483 X-ray weld defect images collected and annotated by oneself were used for architecture optimization through random cropping, rotation, translation, and other data augmentation operations. Finally, a semantic segmentation network for weld defects was automatically constructed at a lower search cost. Experiments show that using the above ideas for semantic segmentation of X-ray weld defects, the final mIoU index reaches 49.23%, which is higher than 45.41% of the artificial design network and 28.86% of the direct use of model transfer. The self-search speed and segmentation effect of the network are significantly improved.

     

/

返回文章
返回