運用卷積神經網路方法分類動物動作;Animal Action Classification using Convolutional Neural Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/73008

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/73008

Title:	運用卷積神經網路方法分類動物動作;Animal Action Classification using Convolutional Neural Networks
Authors:	阮光輝;Huy, Nguyen Quang
Contributors:	資訊工程學系
Keywords:	運用卷積神經網路方法分類動物動作;Animal Action Classification;Deep Learning;Convolutional LSTM;Long-term Recurrent Neural Networks
Date:	2017-01-20
Issue Date:	2017-05-05 17:38:23 (UTC+8)
Publisher:	國立中央大學
Abstract:	在本文中，我們提出了一種用於學習動物活動視頻特徵的新方法。據我們所知，由於動物之相關視頻較少，沒有任何先前的工作在討論此問題。此論文的第一個貢獻在於創建一個全新的動物活動行為數據，分類動物之各種活動行為。此數據之提供來源，不僅為寵物類型作為一些相關作品，而在類似活動中主要是分析之對象為貓和狗。此外，數據自然視頻記錄存在各種天氣場景之條件，如不同之光線。因此，為一種讓任何模型做預測和分類任務之挑戰。我們的第二個貢獻是研究兩個深度學習模型來對我們自己已知數據集進行分類。我們在Torch 7框架中實現兩個空間長短期存儲器架構LRCN和卷積LSTM。我們的設計使網絡變得更方便適應。我們的方法中，訓練我們關於動物行為數據集之架構，在訓練中發現他們具有有學習意義之能力，甚至能套用在一些困難的寵物活動行為之視頻，而網絡在分類任務中能夠獲得相當高的精度約65-70％的結果。我們認為動物行為數據蒐集和深度學習對於更進一步之相關研究探討是非常重要的。;In this dissertation, we proposed a new approach for learning features of animal activity videos. To our best knowledge, there have not any previous work on this problem due to the lack of animal videos. Our first contribution is creating a whole new animal action data which must be difficult to classify in several aspects. The data does not contain only one pet type as some related works, but has two types of cat and dog in similar activities. Furthermore, the data natural videos which are recorded in daily life in variety conditions of light, weather and scene. Thus, our data really challenge any model to do prediction and classification task. Our second contribution is investigating abilities of two Deep Learning models on classifying our own dataset. We implement two spatial Long Short-term Memory architectures LRCN and Convolutional-LSTM in Torch 7 framework. Our design makes adapting the networks easy and convenient. We trained our architectures on animal action dataset and discovered that they have potentials to learn meaning features even of difficult pet videos. The networks get a quite high result on classification task with 65-70% of accuracy. We believe that the animal action dataset and Deep Learning are essential for further studying with more critical requirements.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	392	View/Open

社群 sharing

Loading...