請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88676
標題: | 許超解析度:印刷電路板X光影像之超解析度 HsuSR: Super Resolution for Printed Circuit Board X-Ray Images |
作者: | 許銘真 Ming-Chen Hsu |
指導教授: | 傅楸善 Chiou-Shann Fuh |
關鍵字: | 印刷電路板X光影像,傳統電腦視覺超解析度演算法,許超解析度,雜湊表,CUDA C/C++, Printed Circuit Board,traditional computer vision super-resolution algorithms,HsuSR,hash table,CUDA C/C++, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 印刷電路板PCB (Printed Circuit Board)X光影像在電子產業中是非常重要的一部分,大部分公司會利用PCB的X光影像進行快速的瑕疵檢測,以此確保產品的完整性。為了加快檢測的速度,檢測的過程中會將影像縮小並進行其他任務(例如: 3D錫球重建),最後再放大到原始大小進行瑕疵檢測,由此可知,X光影像品質的好壞是非常重要的。若放大倍率後的影像品質發生失真或是細節消失的狀況,則會讓瑕疵檢測系統辨認出現錯誤。
近幾年來有很多深度學習的方法運用在超解析度上,然而業界公司會接到各式各樣PCB板瑕疵檢測的案子,不可能每次為了新的板子便訓練新的超解析度模型來提升板子超解析度的準確度,而且通常使用深度學習的方法進行超解析時會需要進行大量的運算,所以會花費較久的時間。 為了解決此惡性循環,使用傳統的電腦視覺超解析度演算法較好,但當放大倍率高時,容易產生形狀不正確或是邊緣鋸齒狀等問題發生,因此本論文提出許超解析度演算法,是基於Google在2017所發表的RAISR (Rapid and Accurate Image Super-Resolution) [19]的演算法,初步會利用Bilinear內插法進行放大,並利用雜湊表的方式取得每個像素對應學習好的濾波器並進行卷積計算,以此針對失真的部分進行微調。此外,本篇論文著重於如何將程式進行加速,我們會將程式利用CUDA (Compute Unified Device Architecture) C/C++程式庫運行在GPU (Graphics Processing Unit)上,以此達到我們期望的速度。 The X-ray image of a Printed Circuit Board (PCB) is a crucial part of the electronics industry; as most companies use it for rapid defect detection to ensure product integrity. To speed up the detection process, the images will be zoomed in and subjected to other tasks, such as 3D solder ball reconstruction, during the detection process, and then enlarged back to their original size for defect inspection. Therefore, the quality of X-ray images is crucial. If the image quality is distorted, or details are lost after enlargement, the defect inspection system may make errors. In recent years, there have been many deep-learning methods applied to super-resolution. However, companies in the industry receive various PCB defect inspection projects; it is impossible to train a new super-resolution model every time for a new PCB to improve the accuracy of the super-resolution. Moreover, using deep-learning methods for super-resolution usually requires much computation taking a long time. To break this vicious cycle, traditional computer vision super-resolution algorithms are more suitable. Problems such as incorrect shapes of objects or jagged edges may arise when the magnification ratio is high. Therefore, we propose HsuSR based on RAISR (Rapid and Accurate Image Super-Resolution) [19] algorithm proposed by Google. We will initially use bilinear interpolation for scaling and use a hash table to obtain the learned filters for each pixel and perform convolution to fine-tune the distorted parts. Besides, we focus on acceleration with GPU (Graphics Processing Unit) using the CUDA (Compute Unified Device Architecture) C/C++ library to achieve the expected speed. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/88676 |
DOI: | 10.6342/NTU202302744 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 資訊網路與多媒體研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf | 2.61 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。