RCE-YOLO: receptive-field attention and content-aware feature enhancement for aerospace transparent-component defect detection

Aerospace transparent components are safety-critical optical structures whose surface integrity directly affects visual quality, environmental isolation and maintenance reliability. Single-image detection is particularly difficult when defects are weak, small or elongated and when reflections and illumination changes produce strong background responses. This study develops RCE-YOLO, a compact YOLOv11n-based detector organized around three staged operations: receptive-field-aware feature extraction, content-aware feature reconstruction and multi-scale feature recalibration. RFCBAMConv and C3k2_RFCBAMConv strengthen channel-selective and receptive-field spatial responses; CARAFE replaces fixed top-down interpolation with content-dependent feature reassembly; and efficient multi-scale attention (EMA) recalibrates the final P3, P4 and P5 features before prediction. Experiments use the Aircraft Glass Defect Dataset (AGDD) as a representative aerospace glazing benchmark. The 219 selected single-view source images are split into 197 training and 22 validation images before offline augmentation, yielding 1,379 training and 154 validation images under the current protocol. RCE-YOLO achieves 89.77% precision, 88.29% recall, 93.80% [email protected], 77.27% [email protected] and 71.60% [email protected]:0.95. Relative to YOLOv11n, [email protected] and [email protected] improve by 1.47 and 2.11 percentage points, respectively, with 2.80 M parameters and 7.40 GFLOPs. The eight-configuration ablation includes the three pairwise variants \(A+B\) , \(A+C\) and \(B+C\) ( A = C3k2_RFCBAMConv, B = CARAFE, C = EMA). The \(A+B\) configuration attains the highest [email protected] (93.96%) and [email protected]:0.95 (73.53%), whereas the full \(A+B+C\) model yields the highest [email protected] (77.27%). Pairwise interaction contrasts are 0.00 percentage points at the reported precision for these three mAP metrics; the three-way contrast varies by metric. These point estimates do not demonstrate reproducible synergy or establish that all three stages are necessary. At the class level, contusion [email protected]:0.95 declines by 3.44 percentage points despite a 0.42-point gain at IoU 0.5; the full model therefore does not improve every defect class or IoU criterion. All performance figures are single-run point estimates, without multi-seed standard deviations, confidence intervals or significance tests. Because the 154 augmented validation images originate from only 22 source images, the small numerical differences must not be interpreted as statistically established improvements. No independent external dataset or untouched real-image test set was evaluated; therefore cross-material and cross-acquisition generalization remains unverified.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-15
DOI
https://doi.org/10.1038/s41598-026-80140-w
Primary Topic
Industrial Vision Systems and Defect Detection
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

RCE-YOLO: receptive-field attention and content-aware feature enhancement for aerospace transparent-component defect detection

Zhen He, Yingjie Zhao, Rui Zhang
Scientific Reports
Industrial Vision Systems and Defect Detection
article

RCE-YOLO: receptive-field attention and content-aware feature enhancement for aerospace transparent-component defect detection

Zhen He, Yingjie Zhao, Rui Zhang
article en

Abstract

Aerospace transparent components are safety-critical optical structures whose surface integrity directly affects visual quality, environmental isolation and maintenance reliability. Single-image detection is particularly difficult when defects are weak, small or elongated and when reflections and illumination changes produce strong background responses. This study develops RCE-YOLO, a compact YOLOv11n-based detector organized around three staged operations: receptive-field-aware feature extraction, content-aware feature reconstruction and multi-scale feature recalibration. RFCBAMConv and C3k2_RFCBAMConv strengthen channel-selective and receptive-field spatial responses; CARAFE replaces fixed top-down interpolation with content-dependent feature reassembly; and efficient multi-scale attention (EMA) recalibrates the final P3, P4 and P5 features before prediction. Experiments use the Aircraft Glass Defect Dataset (AGDD) as a representative aerospace glazing benchmark. The 219 selected single-view source images are split into 197 training and 22 validation images before offline augmentation, yielding 1,379 training and 154 validation images under the current protocol. RCE-YOLO achieves 89.77% precision, 88.29% recall, 93.80% [email protected], 77.27% [email protected] and 71.60% [email protected]:0.95. Relative to YOLOv11n, [email protected] and [email protected] improve by 1.47 and 2.11 percentage points, respectively, with 2.80 M parameters and 7.40 GFLOPs. The eight-configuration ablation includes the three pairwise variants \(A+B\) , \(A+C\) and \(B+C\) ( A = C3k2_RFCBAMConv, B = CARAFE, C = EMA). The \(A+B\) configuration attains the highest [email protected] (93.96%) and [email protected]:0.95 (73.53%), whereas the full \(A+B+C\) model yields the highest [email protected] (77.27%). Pairwise interaction contrasts are 0.00 percentage points at the reported precision for these three mAP metrics; the three-way contrast varies by metric. These point estimates do not demonstrate reproducible synergy or establish that all three stages are necessary. At the class level, contusion [email protected]:0.95 declines by 3.44 percentage points despite a 0.42-point gain at IoU 0.5; the full model therefore does not improve every defect class or IoU criterion. All performance figures are single-run point estimates, without multi-seed standard deviations, confidence intervals or significance tests. Because the 154 augmented validation images originate from only 22 source images, the small numerical differences must not be interpreted as statistically established improvements. No independent external dataset or untouched real-image test set was evaluated; therefore cross-material and cross-acquisition generalization remains unverified.

Scientific Reports
Shandong Xiehe University (CN)
Industrial Vision Systems and Defect Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon