高光谱成像
RGB颜色模型
计算机科学
卷积(计算机科学)
特征(语言学)
人工智能
极限(数学)
曲面(拓扑)
集成电路
特征提取
钥匙(锁)
电子线路
失真(音乐)
计算机视觉
频道(广播)
印刷电路板
迭代重建
光谱形状分析
遥感
表面贴装技术
电子工程
质量(理念)
曲面重建
产品(数学)
模式识别(心理学)
图像(数学)
光谱特征
作者
Haijiang liu,Xiaojuan Liao,G Z Chen,Feng Zhao
标识
DOI:10.1088/1361-6501/ae7fa4
摘要
Abstract Flexible printed circuits (FPCs) are critical components in modern electronic products, where reliable surface defect detection is essential to ensure product quality and operational safety. Hyperspectral imaging (HSI) provides rich spectral information for material discrimination, yet its high cost and low efficiency limit large-scale industrial deployment. To address this issue, this paper proposes a novel framework that reconstructs hyperspectral features directly from RGB images without real HSI data samples. The framework employs a Multi-Scale Swin Transformer++ (MST++) for spectral reconstruction, which is integrated with an enhanced YOLOv8s backbone. Key improvements include: (1) group convolution feature extraction to capture spectral correlation; (2) a C2f_IREMA module for effective multi-scale fusion; (3) an SPPF_CPCA block with channel attention to refine feature representation; and (4) a DySample for refined upsampling. The resulting you only look once-based Reconstructed HSI (YOLO-RHSI) model was evaluated on the self-built FPCSD2024 dataset. It achieved an mAP@0.5 of 75.3% and mAP@0.5:0.95 of 43.3%, outperforming YOLOv8s by 5.6% and 3.9%, respectively. These results demonstrate the feasibility of hyperspectral reconstruction as a cost-effective, data-efficient solution for FPC surface defect detection.
科研通智能强力驱动
Strongly Powered by AbleSci AI