透過影像辨識技術完成自動菌落計數App

A Robust and Accurate Automated Colony Counter App Based on Image Processing

簡岳銘(Yueh-Ming Chien)

指導教授 ： 黃乾綱

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2015年)

https://doi.org/10.6342/NTU.2015.01813

摘要

菌落數計算在微生物實驗中是重要的一環，目前的計算方式多為以人工計數。然而面對數量龐大的實驗菌落數，人工計數不僅需花費大量時間、效率低落，錯誤率也會隨計算時間增加而提升。市面上雖已有許多自動計算菌落的儀器，然而價錢昂貴，一般實驗室往往無法負擔。此外，許多相關研究開發出自動計算菌落數的系統，但大多需要固定設備環境以完成計數工作。即便有研究釋出開放原始碼及開放軟體，然而皆未被廣泛採用。有鑑於近年來智慧型手機大量普及化的狀況，以及為了改善上述缺點，本研究欲以手機為計數儀器，開發一套自動計算菌落的App程式，期望以手機平台相對便宜的硬體條件及本研究針對多變攝影條件提出之演算法改善目前菌落計數不便的現況。
本研究提出一種新的影像切割方式，將影像切割成許多小格，並假設格子中大多為背景的情況下，用主成份分析方式，建立背景模型，再進一步以區域成長法找出背景區域，分割出前景與背景。前景中夠符合圓形的區域即視為菌落。接著，將這些區域的RGB histogram及HOG特徵值抽取出來，並依尺寸建立出各個尺寸下的平均菌落模型。最後比對菌落模型與剩下未確定的區域，找出所有菌落並計算個數。
本研究的演算法所計算出來的平均準確率為80%左右，而相關研究中最準確的開放軟體─OpenCFU，在相同條件設定下，其平均準確率為48%，證實本研究的方法在影像品質不佳的情況下，仍有可靠的準確率，進而達到運用手機App自動計算菌落的目的。

關鍵字

菌落計數 ； 影像切割 ； 特徵擷取

並列摘要

Colony counting is an important part of microbiological experiment. At present, colony number is usually counted by manual method. It is time-consuming, inefficient and high error rate when the amount of experimental samples is large. Many automatic colony counting system and machines are developed so far, but they are either expensive or inconvenient to use. Some authors have recently developed automatic colony counting systems, but they require fixed equipment to capture the image. Some authors even release open-source code or software to count colonies. However, none of them is widely adopted. As a consequence, by the popularization of the smart phone, we intend to develop an automatic colony counting app on cellphone to improve it.
We propose a new method of image segmentation to get foreground region. First, split the image into many grids and assume most pixels in each grid belong to background region. Next, use principal component analysis to build background model, which is used by region growing to find all background region and separate foreground from image. We take the foreground regions as colony regions if they are like a circle. Extract RGB histogram and HOG features from colony regions to build different sizes of colony model. Finally, match the remaining region with colony models to get all colony regions and count the number of colonies.
Our approach’s accuracy is 80% and outperforms the best open software of the related research─OpenCFU, which accuracy is 48%. It is proved that our method is robust and accurate to count the colonies automatically on mobile App.

並列關鍵字

colony counting ； image segmentation ； feature extraction

參考文獻

12. 張証凱, 基於智慧型手機自動菌落計數系統, in 工程科學及海洋工程學研究所. 2013, 臺灣大學. p. 1-56.

5. Brugger, S.D., et al., Automated counting of bacterial colony forming units on agar plates. PloS one, 2012. 7(3): p. e33695.

6. Clarke, M.L., et al., Low‐cost, high‐throughput, automated counting of bacterial colonies. Cytometry Part A, 2010. 77(8): p. 790-797.

7. Sieuwerts, S., et al., A simple and fast method for determining colony forming units. Letters in applied microbiology, 2008. 47(4): p. 275-278.

8. Cai, Z., et al., Optimized digital counting colonies of clonogenic assays using ImageJ software and customized macros: comparison with manual counting. International journal of radiation biology, 2011. 87(11): p. 1135-1146.