資料載入中.....
|
請使用永久網址來引用或連結此文件:
http://ir.lib.ncu.edu.tw/handle/987654321/89793
|
| 題名: | 以二胡學習為例之動作識別與糾正的混合分類方法;A Hybrid Classification Method for Action Recognition and Correction – Learning Erhu as An Example |
| 作者: | 佩馬納;Permana, Aditya |
| 貢獻者: | 資訊工程學系 |
| 關鍵詞: | : 動作識別, CNN, 3D-CNN, LSTM, YOLOv3, GCN;Action Recognition, CNN, 3D-CNN, LSTM, YOLOv3, GCN |
| 日期: | 2022-07-25 |
| 上傳時間: | 2022-10-04 11:59:54 (UTC+8) |
| 出版者: | 國立中央大學 |
| 摘要: | 動作識別是深度學習方法的一種實現形式，目前應用於信息技術、體育和藝術
等更廣泛的領域。二胡是起源於中國的一種弦樂器。在演奏這種樂器時，如何正確定
位演奏者的身體並正確握住樂器是有規則的。我們需要一個系統來檢測每位二胡演奏
者的動作來滿足這些需求。因此，本研究將討論使用 3D-CNN、YOLOv3 和 GCN 三種
方法對視頻進行動作識別。 3D-CNN 方法是一種具有 CNN 基礎的方法。 CNN 是一種
常用的圖像處理方法。 3DCNN 已被證明可以有效地從連續視頻幀中捕獲運動信息。為
了提高捕獲每個動作中存儲的每個信息的能力，有必要在 3D-CNN 模型中結合 LSTM
層。 LSTM 是一種高級 RNN，一種順序網絡。它能夠處理 RNN面臨的梯度消失問題。
本研究中使用的另一種具有圖像處理能力的方法是 YOLOv3。 YOLOv3 是一個準確度
水平比較好的目標檢測器，可以實時檢測目標。然後為了最大化 YOLOv3 的性能，本
研究將 YOLOv3 與 GCN 結合起來，這樣我們就可以使用身體關鍵點來幫助 YOLOv3
方法更容易進行分類。 GCN 通過合併圖上局部相鄰周圍節點的幾個特徵來執行空間卷
積。本研究使用 RGB 視頻作為數據集，預處理和特徵提取有三個主要部分。三個主要
部分是身體，二胡桿和弓。為了執行預處理和特徵提取，本研究提出了兩種方法。第
一種方法利用 MaskRCNN 方法對輸入視頻進行分割處理。第二種方法使用身體標誌對
身體片段進行預處理和特徵提取。相比之下，二胡和弓段使用 Hough Lines 演算法。然
後將根據已定義的類別將三個主要部分劃分成幾個部分。此外，對於分類過程，本研
究提出了兩種被使用的深度學習演算法。本研究將所有深度學習方法與傳統的圖像處
理算法方法相結合。這些組合演算法過程將從二胡演奏者展示的每個動作中產生錯誤
消息作為輸出。
;Action recognition is one form of implementation of the deep learning method, which
is currently used in a wider field related to information technology, sports, and the arts. Erhu is
a stringed instrument originating from China. In playing this instrument, there are rules on how
to position the player′s body and hold the instrument correctly. We need a system to detect
every Erhu player′s movement to meet these needs. So that in this study will discuss action
recognition on video using three methods such as 3D-CNN, YOLOv3, and GCN. The 3D-CNN
method is a method that has a CNN base. CNN is a method commonly used to perform image
processing. 3DCNN has been proved effective in capturing motion information from
continuous video frames. To improve the ability to capture every information stored in every
movement, combining an LSTM layer in the 3D-CNN model is necessary. LSTM is an
advanced RNN, a sequential network. It is capable of handling the vanishing gradient problem
faced by RNN. Another method used in this study that has the ability in image processing is
YOLOv3. YOLOv3 is an object detector with a relatively good accuracy level and can detect
objects in real-time. Then to maximize the performance of YOLOv3, this study will combine
YOLOv3 with GCN so that we can use the body key points to help YOLOv3 methods be easier
for classification. GCN performs spatial convolution by merging several features of nodes
around local neighbors on the graph. This research uses RGB video as a dataset, and there are
three main parts in preprocessing and feature extraction. The three main parts are the body, erhu
pole, and bow. To perform preprocessing and feature extraction, this study proposes two
approaches. The first approach uses a segmentation process on the input video by utilizing the
MaskRCNN method. The second approach uses a body landmark to perform preprocessing and
feature extraction on the body segment. In contrast, the erhu and bow segments use the Hough
Lines algorithm. The three main sections will then be divided into several sections according
to the class that has been defined. Furthermore, for the classification process, this study
proposes two algorithms to be used, namely, deep learning. This study will combine all deep
learning methods with traditional image processing algorithm methods. These combination
algorithm processes will produce an error message output from every movement displayed by
the erhu player. |
| 顯示於類別: | [資訊工程研究所] 博碩士論文
|
文件中的檔案:
| 檔案 |
描述 |
大小 | 格式 | 瀏覽次數 |
|---|
| index.html | | 0Kb | HTML | 21 | 檢視/開啟 |
|
在NCUIR中所有的資料項目都受到原著作權保護.
|
