Skip to content
Conference

Shallow feature enhanced YOLO for microcrack detection in aero-engine blade images

Oct 2026 · Conference on Advanced Algorithms and Signal Image Processing · Vol 14370, pp. 143700B - 143700B-6 · 0 citations · 10 references
Engineering

Abstract

Aero-engine blade micro-cracks are safety-critical defects that are difficult to detect in borescope images because of their small scale, elongated morphology, weak edge contrast, and complex background interference. To improve the detection performance of micro-cracks, this paper proposes a shallow-feature-enhanced YOLO11 detection method for aero-engine blade defect inspection. The proposed framework introduces P1/P2 high-resolution shallow detection branches to preserve fine-grained structural information, including crack endpoints, weak edges, and local texture cues, which are easily degraded during deep network down-sampling. In addition, Focal Loss is incorporated to alleviate the imbalance between abundant background samples and sparse defect samples, enabling the model to focus more effectively on low-confidence and hard crack samples. Experiments were conducted on a real aero-engine borescope image dataset containing 1,171 images and 386 annotated defect instances, including cracks and gouges. The ablation results show that the proposed method improves the F1 score from 0.8751 to 0.9303 and increases recall from 0.8598 to 0.9701 compared with the YOLO11 baseline. Although a slight decrease in precision and localization-related metrics is observed after introducing Focal Loss, the substantial improvement in recall indicates that the proposed method effectively reduces missed detections, which is particularly important for safety-critical aero-engine blade inspection. The results demonstrate that combining shallow high-resolution feature enhancement with hard-sample re-weighting provides an effective solution for micro-crack detection in complex borescope scenarios.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.