Key points are not available for this paper at this time.
Detecting damage precursors in metal additively manufactured (AM) components poses a significant challenge for industrial quality assurance, especially on highly reflective surfaces that severely degrade the performance of conventional optical inspection methods. To address the persistent issues of small-defect recognition and environmental robustness, this study proposes an enhanced SOutlook-YOLOv11 framework that synergistically integrates polarized imaging with an optimized deep-learning architecture. The proposed model advances the baseline YOLOv11 through three key improvements: (1) introduction of the C3K2-Outlook module, which embeds outlook attention across the network to strengthen feature representation; (2) insertion of a spatial adaptive feature modulation (SAFM) module into the neck to improve multi-scale feature fusion and integration; and (3) replacement of the CIoU loss with an Inner-CIoU loss function, thereby enhancing bounding box localization accuracy and detection confidence for subtle damage precursors. Experimental results demonstrate that the proposed model substantially outperforms the baseline YOLOv11, achieving increases of 4.2% in precision, 0.8% in recall, and 1.9% in mAP@50. Notably, the recall reaches 0.937, confirming the model’s high reliability in identifying damage precursors. This work presents a novel strategy, to our knowledge, that bridges optical physics with deep learning, offering a more reliable and intelligent solution for quality inspection in high-precision additive manufacturing systems.
Peng et al. (Mon,) studied this question.