使用深度學習之X光影像超解析度

Super-Resolution with Deep Learning for X-ray Images

蔡婷瑀(Ting-Yu Tsai)

指導教授 ： 傅楸善

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202101004

摘要

本論文提出深度學習的模型來進行光學影像超解析度。因為電荷耦合裝置CCD (Charge-Coupled Device) 影像感測器在製造過程上的改進，加上數位影像處理技術快速的發展，有不少的測量系統直接採用感光耦合元件做為快速的檢測影像裝置。而無論光學系統是用來進行測量或查看，光學影像的品質優劣是最重要的，攸關整個光電成像系統的良劣。也因此若光學影像超解析度後的品質失真或細節變模糊，則會無法檢測到正確的瑕疵。傳統電腦視覺演算法在放大倍率高的時候會產生邊緣鋸齒狀或細節模糊的情況，因此我們提出深度學習的模型，來還原細節。近年來，生成對抗網路 (Generative Adversarial Network, GAN) 在影像的生成、合成、辨識、修復等等方面具有出色的性能。我們設計一套超解析度流程，參考幾種最先進模型的優點後，提出一個新的模型：蔡網路 (TsaiNet)，在GAN的基本架構上加了Wasserstein距離 (Wasserstein distance)以及特徵判別器 (feature discriminator)，產生更有意義的影像細節。與數種經典的GAN模型和Python 內插法相比，我們的超解析度方法具有最佳的性能和快速的測試速度。

關鍵字

Optical Image Detection ； Super-Resolution ； TsaiNet ； Generative Adversarial Network ； Artificial Intelligence ； Image Processing

並列摘要

In this thesis, we propose a deep learning model for super-resolution optical images. Because of improvement in manufacturing process of CCD (Charge-Coupled Device) sensors and the rapid development of digital image processing technology, many measurement systems directly use CCD as fast detection image device. Whether optical system has been measured or viewed, optical image quality is the most important, assisting the analysis of the quality of the entire optical imaging system. If optical image quality after super-resolution produces distortion or blurred image details, we can not inspect defects in the images. Moreover, traditional computer vision algorithm will produce jagged or blurred details with high magnification.

We have proposed a deep learning model to restore meaningful details. Recently, Generative Adversarial Network (GAN) has excellent performance in image generation, synthesis, identification, restoration, and so on. We designed a super-resolution process, referring to the advantages of several state-of-the-art models, proposed a new model TsaiNet.

TsaiNet adds Wasserstein distance and feature discriminator to basic structure of GAN to produce more meaningful image details. Compared with several classic GAN models and Python interpolation methods, our super-resolution model, TsaiNet, has the best performance and fast testing speed.

並列關鍵字

光學影像檢測 ； 超解析度 ； 蔡網路 ； 生成對抗網路 ； 人工智慧 ； 影像處理

參考文獻

M. S. Alam, J. G. Bognar, R. C. Hardie, and B. J. Yasuda, “Infrared Image Registration and High-Resolution Reconstruction Using Multiple Translationally Shifted Aliased Video Frames,” IEEE Transactions on Instrumentation and Measurement, Vol. 49, No. 5, pp. 915-923, 2000.

C. Dong, C. C. Loy, and X. Tang, “Accelerating the Super-Resolution Convolutional Neural Network,” Proceedings of European Conference on Computer Vision, Amsterdam, The Netherlands, pp. 391-407, 2016.

C. Dong, C. C. Loy, K. He, and X. Tang, “Image Super-Resolution Using Deep Convolutional Networks,” Proceedings of IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 38, No. 2, pp. 295–307, 2016.

M. Elad and Y. Hel-Or, “A Fast Super-Resolution Reconstruction Algorithm for Pure Translational Motion and Common Space-Invariant Blur,” IEEE Transactions on Image Processing, Vol. 10, No. 8, pp. 1187-1193, 2001.

Google AI Blog, “The Google Brain Team — Looking Back on 2017 (Part 1 of 2),” https://ai.googleblog.com/2018/01/the-google-brain-team-looking-back-on.html, 2018.