Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/79221
Title: | 應用影像前處理以提升人物動作辨識網路之效能 Application of Image Preprocessing to Improve Human Action Recognition |
Other Titles: | Application of Image Preprocessing to Improve Human Action Recognition |
Authors: | 林佩宜 Pei-Yi Lin |
Advisor: | 丁肇隆 Chao-Lung Ting |
Keyword: | 人物動作辨識,卷積神經網路,深度學習,影像處理,壓縮影片格式, human action recognition,convolutional neural network,deep learning,image processing,compressed video, |
Publication Year : | 2022 |
Degree: | 碩士 |
Abstract: | 近年來,深度學習技術高度發展,以及計算能力的進步,圖像中的物件辨識及分類的功能進步顯著,隨著影片資訊量的成長,對影片中人物進行動作辨識也逐漸受到重視,但是影片中的動作事件必須透過連續的畫面才能正確判斷,而其中重複的畫面包含許多冗餘資訊,因此[1]提出使用壓縮資訊的方式,從壓縮影片中直接提取運動向量及殘差資訊,作為訓練所需之輸入資訊,以減少訓練成本,而本研究基於CoViAR的方法,並加以改進。實驗結果證實,若以更多的動作資訊取代色彩的變化,並將運動歷史圖應用於殘差影像中,能夠提升辨識準確率。另外,將單幀影像及運動向量資訊融合,作為一個模型的輸入資訊,並且將殘差影像的三通道模型縮減為單通道,可以減少模型訓練參數,以更低的計算成本達到相當程度的辨識效果。 In recent years, object recognition and classification in images have advanced significantly due to deep learning technology which has been highly developed and the computing power has advanced. Therefore, [1] proposed the method of compressed information to extract motion vectors and residual information directly from compressed videos as input information for training to reduce training costs. This study is based on the method of CoViAR and improved. The experimental results show that the recognition accuracy can be improved if more motion information is used instead of color variation, and the motion history image technology is applied to the residual images. In addition, combining I-frame image and motion vector information as input of one model, and reducing the three-channel model of the residual image to a single channel can reduce the model parameters and achieve a considerable recognition performance with lower computational cost. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/79221 |
DOI: | 10.6342/NTU202210002 |
Fulltext Rights: | 同意授權(全球公開) |
metadata.dc.date.embargo-lift: | 2023-11-09 |
Appears in Collections: | 工程科學及海洋工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
U0001-1028221026495102.pdf | 2.8 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.