超頻譜生醫影像系統之巨量資料處理與雲端操控之研究

Abstract

此論文提出一種新的超頻譜生醫影像平台，此平台以嵌入式繼光顯微超頻影像系統 (ERL-HIS)作為代表超頻譜生醫影像的載具，並整合了雲端控制以及即時壓縮於其中。整合於ERL-HIS的雲端控制能夠透過連結機台的伺服器與遠端用戶做連結，進而操控ERL-HIS的掃描動作並同時做資料接收。 ERL-HIS內的MFC伺服器與Android 用戶端彼此互相傳送訊息，這些訊息依賴設計好的命令架構做儀器操控以及資料傳輸。除此之外，此種超頻譜生醫影像應用於巨量資料上的概念也將被討論於此論文，主要討論所提出的壓縮方法與Hadoop平行運算做結合之應用細節，以及一種壓縮率略差但是可以快速解壓縮出想要的特定頻譜資訊之方法。 另外，此論文也著重於ERL-HIS影像的壓縮，分為充滿雜訊的原影像之壓縮與去雜訊後的影像壓縮，所提出的壓縮法主要利用鄰近像素的整條光譜來預測尚未被預測的光譜之趨勢，該種方法與著名的LUT法以及3D-CALIC的預測路徑(相鄰頻帶間的預測)是完全不同的。 對於充滿雜訊的原影像,所提出的方法在資料殘餘之熵值上比起LUT與3D-CALIC還要少1bpp以上，因此可以達到更好的壓縮率。而對於去雜訊後的影像，所提出的方法在熵值上比起LUT方法少約0.5bpp，比起3D-CALIC則少約0.3bpp，因此在壓縮上仍然可以稍微得到較佳的壓縮率。在執行時間上此三者的比例為: 1 : 47.58 : 4.25 (LUT:3D-CALIC:所提出的方法(去雜訊版本)) 1 : 47.58 : 13.83 (LUT:3D-CALIC:所提出的方法(雜訊版本)) 可以看出，在壓縮率上此論文所提出的方法可以達到最佳的壓縮率，且有著比起3D-CALIC快上數倍的執行速度。

In this thesis, a novel platform for an embedded relay lens hyperspectral imaging system (ERL-HIS) is presented. This type of platform integrates cloud control and real-time compression on ERL-HIS. Cloud control is designed to operate the ERL-HIS remotely through a cloud server and users can manipulate the ERL-HIS during scanning for image transformation. The details of how to make commands from android clients to the ERL-HIS using an MFC server are discussed. Additionally, a concept for biomedical big data is discussed. We delineate the utilization of the proposed compression method into Hadoop for parallel computing. Also, a modified compression designed for quick access decoding path is discussed.

A novel compression method for noisy signals from an ERL-HIS and de-noised signals are proposed. The proposed method predicts pixels using the tendency of neighbors’ spectrums while the other two methods, the LUT method and the 3D-CALIC method, predict pixels band by band. These proposed compression methods are compared to the LUT method and the 3D-CALIC method with regard to compressing an AVIRIS hyperspectral image. The proposed method performs with a better compression ratio (CR) than LUT and 3D-CALIC with regard to noisy signals from the ERL-HIS and its residuals cost about 1bpp less other methods. When compressing de-noised signals, entropy of the residuals produced from proposed method is less than LUT’s about 0.5bpp and is less than 3D-CALIC’s about 0.1 bpp so it still has better CR than other methods. The average runtime proportion of these methods are as follows: 1 : 47.58 : 4.25 (LUT:3D-CALIC:proposed method(de-noised version)) 1 : 47.58 : 13.83 (LUT:3D-CALIC: proposed method (noisy version))

Thus, it is suggested for compressing ERL-HIS biomedical images if a high CR is required.