利用類神經網路於浮游藻類自動影像辨識分類之研究

Identification of Algae with Pattern Recognition by Artificial Neural Network

楊格(Ke Yang)

指導教授 ： 吳先琪

國立臺灣大學/工學院/環境工程學研究所/碩士(2012年)

https://doi.org/10.6342/NTU.2012.01551

摘要

水庫優養化是造成水質惡化之主要原因之一，本研究欲建立一節省時間及人
力之自動化藻類辨識系統，以得知自然水樣中的藻種。研究分為硬體設計及軟體
撰寫兩部分。
硬體部分包含進樣後之水樣自動濃縮裝置、流動式樣品觀測槽、顯微鏡及
CCD。自動濃縮裝置是以切向流原理設計，配合蠕動幫浦進樣可達到自不須反沖
洗，自動濃縮水樣之效果。流動式樣品藏是參考流式細胞儀原理，製作一使水樣
可流動之淺槽，配合顯微鏡及高速攝影CCD，可連續隨機拍攝水樣之影像，以
節省製作玻片之材料與時間。
軟體部分分為影像前處理及影像辨識兩部分，本研究以Matlab 語言撰寫自
動藻類影像辨識之軟體，提取藻類訓練樣本影像之型態與色彩特徵值，訓練倒傳
遞類神經網路，經學習後的倒傳遞類神經網路可對未知的藻類影像做自動辨識。
此系統之系統對於人工培養藻之辨識率為92%、其中單胞藻辨識率為87%、
藍綠藻辨識率為87%、微囊藻辨識率為93%、柱珠藻辨識率為93%，而雜質辨
識率為98%。而對於自然水體中的平裂藻辨識率為70%、扭曲單殼縫藻為50%、
星鼓藻為73%、盤星藻為80%。
結果顯示人工培養藻類的辨識率高於自然水體藻類，未來欲應用至自然水體
藻類監測時，若能對於特定水體事先輸入訓練樣本影像，將類神經網路依不同常
見藻類做學習和調整，則可成為監測此水體藻類生長情況的有效工具。

關鍵字

藻類 ； 影像辨識 ； 倒傳遞類神經網路

並列摘要

Eutrophication is the most common cause for the deterioration of reservoir water
quality in Taiwan. Identification of algal species and estimation of the abundance are
necessary for the warming and managing the situation of eutrophication. The aim of
this research is to establish an automatic algae recognition system which is not only to
recognize the species of phytoplankton in natural water sample but also to reduce the
time cost and labor. This research could be divided into two parts, the hardware
development and the software coding.
The hardware includes a sample injector followed by the automatic condensation
equipment, the flowing cell, a microscope and a CCD. The designing of the automatic
condensation equipment was based on the tangential flow filtration principle. The
water sample was driven by Ismatec Peristaltic pump, into the automatic condensation
equipment which operated smoothly without backwash. According the concept of to
Flow Cytometry, this research devised a shallow flow trough cell called flowing cell.
The high speed CCD would capture the digital image continuously while water
sample passing through this flowing cell. This approach reduced the material and time
cost of making glass coverslips.
The software coding was composed of image pre-treatment and image
recognition. We wrote an automatic algae recognition program by Matlab language
and trained the Back-Propagation Neural Network model by inputing extracted
configuration features and color features of the training pictures. The trained
Back-Propagation Neural Network is able to recognize unknown algae cells and
colonies.
The recognition accuracy for a mixture of four artificial cultivated algal species
was 87% for Chlamydomonas, 87% for Cyanobacteri, 93% for Melosira granulate
IV
and 93% for Microcystaceae. In addition, the system recognition accuracy was 70%.
For Merismopedia, 50% for Monoraphidium Contortum, 73% for Staurastrum and
% for Pediastrum refer in a natural water sample.
As a result, the system recognition accuracy for artificial cultivated algae species
was higher. If we want to apply this monitoring system to natural water body in the
future, we should input specific algae images in the water body to train the Neural
Network model. The Back-Propagation Neural Network would self-adjust and
self-learn. To develop a effective tool for monitoring phytoplanktons in natural
waters.

並列關鍵字

phytoplankton ； pattern recognition ； Back-Propagation Neural Network

參考文獻

李玫憶 (2005)，細胞次結構影像辨識系統，中原大學醫學工程系，碩士論文。

張仁明、林達德，(1999) 類神經網路與貝氏分類法應用於影像分割之比較研究，農業機械學刊，81, 3, 61-74。

張逸中、蘇家弘 (2007) 單細胞藻類之動態影像辨識，電腦與通訊應用，1, 71-80。

郭彥甫、林達德、王嘉瑞. (2004)，應用影像處理技術輔助植物葉片之幾何模擬，農業機械學刊，13.1-18。

曾麗文 (2010)，貝氏分類器於醫學數據資料探勘之研究，國立成功大學工業與資訊管理學系專班，碩士論文。