PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
April 5, 2026AIP Advances2 citationsOpen Access

SGCDT-YOLO: A multi-scale feature fusion network with content-aware selective mechanisms for wood surface defect detection

View Full Paper
JWJun WuAZAiguo ZhangCDChao Deng

Key Points

  • To improve wood surface defect detection by addressing multi-scale challenges and enhancing recognition accuracy.
  • Developed SGCDT-YOLO, an improved YOLOv11n algorithm.
  • Integrated Selective Weighted Spatial Reduction for content-aware downsampling.
  • Implemented Guided Expansion and Dimensional-Oriented Selective Fusion Enhancement Network for feature interaction.
  • Designed Dual-Stage Gradient Receptive Field Aggregation to extend perceptual range with low computational overhead.
  • Incorporated triplet attention modules to enhance multi-morphology defect recognition.
  • Achieved 78.12% mAP@0.5 and 40.42% mAP@0.5:0.95, improvements of 13.41% and 5.43% over the baseline.
  • Increased precision to 76.62% (+10.25%), recall to 68.52% (+6.76%), and F1-score to 72.34% (+8.36%).
  • Reduced the number of parameters by 10.47% compared to the baseline model.
  • Outperformed mainstream algorithms by 12.22–16.02% mAP@0.5 in comparative tests.

Abstract

Wood surface defect detection is plagued by challenges in multi-scale feature fusion, small object recognition, and other aspects. To address these issues, this paper proposes SGCDT-YOLO, an improved YOLOv11n algorithm integrated with the Selective Weighted Spatial Reduction (SWSR), Guided Expansion and Dimensional-Oriented Selective Fusion Enhancement Network (GEDOSFEN), Dual-Stage Gradient Receptive Field Aggregation (DSGRFA), and triplet attention modules. A dual-branch SWSR module is designed to enable content-aware downsampling, preserving discriminative information in critical defect regions while suppressing irrelevant areas. GEDOSFEN is proposed to build bidirectional diffusion paths via intermediate-scale features, facilitating effective interaction between deep semantic information and shallow texture details across detection scales. Inspired by dilated-wise residual and dilated re-param block, the DSGRFA module is designed to decompose receptive field expansion into regional feature construction and semantic filtering, extending the perceptual range with low computational overhead. Triplet attention is incorporated to establish cross-dimensional correlations among channel, height, and width, enhancing the unified perception of multi-morphology defects. Experimental results show that SGCDT-YOLO achieves 78.12% mAP@0.5 and 40.42% mAP@0.5:0.95, representing improvements of 13.41% and 5.43% over the baseline, respectively, with precision at 76.62% (+10.25%), recall at 68.52% (+6.76%), and F1-score at 72.34% (+8.36%).Meanwhile, the number of parameters is reduced by 10.47% compared with the baseline model. Comparative tests reveal a 12.22–16.02% mAP@0.5 improvement over mainstream algorithms. Visualization analyses of inference results, heatmaps, and feature maps confirm the algorithm’s high accuracy in defect localization and classification.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Wu et al. (2026) studied this question.

synapsesocial.com/papers/69d1fcc0a79560c99a0a2619https://doi.org/10.1063/5.0328459
Ask AI
Helpful
Bookmark
Share
View Full Paper