- 學位論文
果蠅影像序列的行為標註系統及機器學習辨識
Behavior Annotation System for Drosophila Image Sequence and Behavior Recognition Using Machine Learning
摘要
由於果蠅與人類的基因有許多相似的地方，透過觀測基因改變後果蠅的求偶行為可以了解哪些基因會影響人類的行為。對果蠅求偶動作的進行影像分析，過去是使用人工方式去測量並記錄，導致生物學家們需要花費大量時間觀察及記錄，此方法效率不高且容易出錯。為了降低觀察所耗費的時間並加快分析行為數據，本篇論文開發一個能讓生物學家輕易掌握的果蠅影像序列的行為標註系統，使生物學家在研究及分析果蠅求偶行為能更有效率。後續並利用生物學家對影片中的行為標記，再配合自動影像物件切割程式，可以切割出用於後續機器學習的果蠅行為圖像。 本論文建立機器學習是以DIGITS作為建立機器學習的平台且以Caffe當作機器學習的框架，我們使用AlexNet及GoogLeNet的網路架構來訓練模型。本論文的辨識行為種類共分為Singing、Tapping、Attempting三種，其中Singing與Tapping選用的圖像張數為7602張，Attempting選用的圖像張數為5184張。我們在DIGITS上的使用微調後的AlexNet模型驗證準確率為99.79%、GoogleNet的模型驗證準確率為99.88%。為了驗證微調後的AlexNet及GoogLeNet的模型可靠度，我們在Singing、Tapping、Attempting各選用180張與訓練資料集相同影片產生出的圖像來做測試，經過測試後我們得到AlexNet的準確率為75.9%、GoogLeNet的準確率為93.3%，接著使用Singing、Tapping、Attempting各200張與訓練資料集不同的影片產生出的圖像做測試，得到AlexNet的準確率為62.7%、GoogLeNet的準確率為60%。
並列摘要
Since the genes of Drosophila and humans have many similarities, by observing the courtship behavior of the flies that result from genetic changes, we can understand which genes affect human behavior. In the past, the image analysis of the courtship action of fruit flies used manual methods to measure and record, which caused biologists to spend numerous hours observing and recording. This method is inefficient and error-prone. In order to reduce the time spent on observation and speed up the analysis of behavioral data, this thesis has developed a behavioral labeling system for fruit fly image sequences, which can be easily mastered by biologists so can study and analyze Drosophila courtship behavior more effectively. Subsequent use of biologists’ marking of behavior in the film with the automatic image object segmentation program, can cut out the fruit fly behavior image for subsequent machine learning. In this thesis, machine learning is established by using DIGITS as the machine learning platform and Caffe as the machine learning framework. We also use the network architecture of AlexNet and GoogLeNet to train the model. There are three types of identification behaviors in this thesis: Singing, Tapping, and Attempting. The number of images labeled as Singing and Tapping is 7602 respectively, and the number of images labeled as Attempting is 5184. We used the fine-tuned AlexNet model and the training accuracy rate on DIGITS was 99.79%, on the other hand, the GoogleNet model training accuracy rate was 99.88%. In order to verify the reliability of the fine-tuned AlexNet and GoogLeNet models, we selected 180 images for Singing, Tapping, and Attempting respectively generated from the same videos as the testing data set. After testing, we got the accuracy of AlexNet to be 75.9%, and GoogLeNet's accuracy rate was 93.3%. Then, we selected 200 images for Singing, Tapping, and Attempting respectively generated by different videos as the testing data set. We got AlexNet accuracy rate of 62.7%, and GoogLeNet accuracy rate of 60%.
參考文獻
延伸閱讀
- 沈彥廷（2014）。果蠅種類影像辨識及群組互動監測研究〔碩士論文，國立清華大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0016-2912201413503792
- Chiang, M. H. (2014). 利用基因表現交集來製造果蠅神經網路的操控工具 [master's thesis, National Tsing Hua University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0016-2912201413535048
- 黃建霖（2012）。果蠅基因交互作用之視覺化與網絡分析平台建置〔碩士論文，國立清華大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0016-2002201315224634
- Chen, Y. T. (2018). 建立果蠅過動症行為模式參數 [master's thesis, National Tsing Hua University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0016-1803201914420501
- 賴威成（2010）。Development and Application of Multi-function Drosophila Behavior Chip〔碩士論文，國立清華大學〕。華藝線上圖書館。https://doi.org/10.6843/NTHU.2010.00313