陳超解析度：利用超解析度增強印刷電路板X光影像

Abstract

隨著電子產品高度精密化，印刷電路板（Printed Circuit Board, PCB）設計日趨複雜，對於瑕疵檢測的精度與效率提出更高要求。X光自動檢測（Automated X-ray Inspection, AXI）雖能穿透元件結構，協助發現隱藏瑕疵，但其所獲得之影像常因硬體限制而解析度不足，進而影響後續瑕疵辨識準確性。因此，本論文針對單張X光影像解析度提升問題，提出一種輕量級深度學習超解析度演算法ChenSR，以提升影像細節還原能力，同時保有高運算效率。

本論文所提出之ChenSR架構為改良式的Transformer模型，結合Efficient Hybrid Transformer Blocks（EHTB）與MobileNetV2中的inverted residual設計，有效降低模型參數量與記憶體需求，並透過通道注意力與多層殘差連接保持重建品質。與其他主流深度學習超解析度方法相比，ChenSR展現出極具競爭力的推論速度，使其成為目前深度學習架構中運算效率表現相當優異的選擇之一，特別適合部署於即時或邊緣裝置之工業應用場景。

在實驗設計上，本研究與現有代表性方法 （Bicubic、HsuSR、DRCT與HMANet） 進行比較。實驗結果顯示，ChenSR在影像品質（PSNR與SSIM）、結構誤差評估（Diff）及執行效率（計算時間）方面皆展現優異之平衡表現，同時在0.03秒內完成單張影像4倍放大，滿足即時檢測應用需求。

With the increasing complexity and miniaturization of electronic products, the design of Printed Circuit Boards (PCBs) has become more intricate, raising higher demands for both the accuracy and efficiency of defect inspection. Although Automated X-ray Inspection (AXI) can penetrate component structures to detect hidden defects, the acquired images often suffer from insufficient resolution due to hardware limitations, which negatively impacts the accuracy of subsequent defect recognition. To address the challenge of enhancing single X-ray image resolution, this thesis proposes a lightweight deep learning-based super-resolution algorithm, ChenSR, which aims to recover fine image details while maintaining high computational efficiency.

Our proposed ChenSR architecture is a modified Transformer-based model, incorporating Efficient Hybrid Transformer Blocks (EHTB) and the inverted residual structure from MobileNetV2. This design effectively reduces model parameters and memory usage while maintaining reconstruction quality through channel attention and multi-level residual connections. Compared with other mainstream deep learning super-resolution methods, ChenSR exhibits highly competitive inference speed, making it one of the most efficient frameworks currently available. Its lightweight and efficient design makes it particularly suitable for real-time or edge-device deployment in industrial applications.

In the experimental evaluation, ChenSR is compared with several representative methods, including Bicubic, HsuSR, DRCT, and HMANet. Results demonstrate that ChenSR achieves a strong balance between image quality (measured by PSNR and SSIM), structural error (Diff), and inference time. Notably, it completes 4× image enlargement within 0.03 seconds per image, fulfilling the requirements for real-time defect inspection.