Scholarlink

<p>Accurate and efficient detection of aircraft surface cracks is crucial for aviation safety but remains challenging due to the high computational demands of deep segmentation models. This study proposed EKD-DeepLab, a lightweight semantic segmentation framework that integrates dual-teacher hybrid knowledge distillation with explicit edge enhancement. Two complementary teacher networks—PSPNet for global context learning and DeepLabV3+ for boundary-aware segmentation—jointly guided a compact DeepLabV3+ student model through multi-level distillation, including pixel-wise, pair-wise, and holistic adversarial losses. To further enhance spatial representation, a Laplacian–Gabor Fusion Module (LGFM) strengthened edge and texture perception, while a 2D Quantized-inspired Compact Pyramid (2D-QCP) aggregated multiscale shallow features to preserve fine structural details. Evaluated on the Aircraft Defect (AD) dataset, EKD-DeepLab achieved 98.28% pixel accuracy, 90.12% F1-score, and 90.33% mean Intersection over Union, while reducing parameters by more than 50% and reaching 83.4 FPS on 512 × 512 inputs. These results demonstrated that EKD-DeepLab effectively balanced segmentation precision and computational efficiency, providing a practical and deployable solution for real-time aircraft defect inspection and other lightweight industrial visual inspection tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-teacher knowledge distilled DeepLabV3+ with enhanced edge features for efficient airplane surface crack segmentation

  • 期刊论文

  • 发表日期：

    2025年11月29日

查看全文
  • Dacheng Wang,
  • Zhexin Wang

摘要

Accurate and efficient detection of aircraft surface cracks is crucial for aviation safety but remains challenging due to the high computational demands of deep segmentation models. This study proposed EKD-DeepLab, a lightweight semantic segmentation framework that integrates dual-teacher hybrid knowledge distillation with explicit edge enhancement. Two complementary teacher networks—PSPNet for global context learning and DeepLabV3+ for boundary-aware segmentation—jointly guided a compact DeepLabV3+ student model through multi-level distillation, including pixel-wise, pair-wise, and holistic adversarial losses. To further enhance spatial representation, a Laplacian–Gabor Fusion Module (LGFM) strengthened edge and texture perception, while a 2D Quantized-inspired Compact Pyramid (2D-QCP) aggregated multiscale shallow features to preserve fine structural details. Evaluated on the Aircraft Defect (AD) dataset, EKD-DeepLab achieved 98.28% pixel accuracy, 90.12% F1-score, and 90.33% mean Intersection over Union, while reducing parameters by more than 50% and reaching 83.4 FPS on 512 × 512 inputs. These results demonstrated that EKD-DeepLab effectively balanced segmentation precision and computational efficiency, providing a practical and deployable solution for real-time aircraft defect inspection and other lightweight industrial visual inspection tasks.