.png){kind=icon}
.png){kind=icon}
Return
A multi-scale feature reconstruction defect detection method guided by global semantics
.png){kind=icon}
PRE
AI
DOI：10.1088/2631-8695/ae698b.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The safe and reliable operation of aircraft engines is a crucial prerequisite for ensuring flight safety. Aircraft engine turbine blades operate under extreme conditions and are prone to defects such as wear and cracks, posing a serious threat to flight safety. Therefore, efficient and reliable non-destructive testing of blade defects holds significant engineering importance. Traditional non-destructive testing methods are labor-intensive and prone to errors. In contrast, deep learning approaches based on computer vision demonstrate strong potential for defect detection. However, the existing unsupervised methods are insufficiently robust in complex scenarios due to the scarcity of samples and domain offset. Therefore, this paper proposes an unsupervised defect detection method-global semantic-guided multi-scale reconstruction (GSMR), combining global semantic constraints and multi-scale feature reconstruction. First, the global semantic embedding module extracts multi-scale features with strong cross-domain generalization ability and builds a semantically consistent reference baseline. Subsequently, the mask-guided local reconstruction module performs high-quality feature recovery through random masking and an improved multi-scale feature pyramid network decoder, enhancing sensitivity to anomalous patterns. Finally, a cross-scale residual measurement module is introduced to compare multi-scale differences between reference and reconstructed features. Adaptive Gaussian smoothing is then applied to generate robust anomaly maps. Experimental results show that GSMR outperforms existing methods on the AeBAD-S dataset, achieving an average sample-level AUROC of 88.8% and an average pixel-level per-region-overlap of 90.2%. Additionally, the method exhibits superior performance on the MVTec AD dataset. This validates the effectiveness of the proposed approach for detecting blade defects in complex backgrounds and across different domains.
Keywords:
defect detection of aircraft engine blades
domain shift
global semantic feature embedding
multi-scale feature reconstruction
Journal
E
IF:
1.6
Papers:
2.1K
Citations:
0