基於資料擴增與深度學習的改善磁控膠囊內視鏡辨識腸道之研究

Abstract

近年來，在科技的不斷進步之下，電腦視覺將直接應用在各行各業中。於此同時，由於高效能圖形處理器 (Graphics Processing Unit，GPU) 的效能大幅進步，也使得深度學習的相關研究相當熱門。憑藉良好的訓練模型與充足的訓練數據，深度學習將電腦視覺之研究推進至一個新紀元。 在本論文中，我們將深度學習應用於辨識腸腔並取得其位置，達到磁控大腸膠囊內視鏡自動導航之方法。使用了最新的即時目標偵測深度學習模型－YOLO (You Only Look Once)。我們使用KVASIR之腸道圖像資料集來驗證YOLO模型之腸腔偵測效果，但是KVASIR之腸道圖像的數量仍然很少。 由於腸道圖像的不足，本文目的是使用資料擴增技術使KVASIR之腸道圖像增加，以便將其用於YOLO模型訓練以提高檢測效果。我們使用兩種擴增方式：傳統之資料擴增與生成對抗網路 (Generative Adversarial Networks, GAN) 。傳統之資料擴增使用五種擴增方式：圖像平移、圖像旋轉、圖像縮放、圖像加入高斯雜訊與圖像加入動態模糊；而生成對抗網路使用Cycle-GAN將模擬腸道圖像轉換成KVASIR之腸道圖像。

In recent years, with the continuous progress of science and technology, computer vision will be directly applied in all walks of life. At the same time, due to the efficiency of highly efficient graphics processing unit (GPU) greatly improved, the related research of deep learning is quite popular. With good training models and sufficient training data, deep learning advances the research of computer vision into a new era. In this paper, we apply deep learning technology to identify the lumen and obtain its position, so as to achieve the novel navigation for magnetic field control endoscope. We use the state-of-the-art, real-time object detection deep learning model - YOLO (You Only Look Once). KVASIR dataset was used to evaluate the performance of lumen detection, but images of the KVASIR dataset are still few. Due to insufficient intestinal images, this thesis aims at using data augmentation technology to increase the intestinal images of KVASIR dataset, so as to apply it to YOLO model training to improve the detection effect. We used two data augmentation methods: traditional data augmentation and Generative Adversarial Networks (GAN). Traditional data augmentation methods include image translation, image rotation, image scaling, image adding Gaussian noise and image adding motion blur; Cycle-GAN was used to transform the simulated intestinal images into the real intestinal images of KVASIR dataset.