Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/76612
Title: | 應用於未校正魚眼影像上的人臉偵測模型 FACE DETECTION IN THE FISHEYE IMAGE DOMAIN |
Authors: | Cheng-Yun Yang 楊承運 |
Advisor: | 陳宏銘(Homer H. Chen) |
Keyword: | 魚眼影像,魚眼扭曲,人臉偵測,深度學習,邊緣運算, Fisheye image domain,Fisheye distortion,Face detection,Deep learning,Edge computing, |
Publication Year : | 2020 |
Degree: | 碩士 |
Abstract: | 在魚眼影像上與一般影像上進行人臉偵測最大的不同點在於,在魚眼影像邊緣區域的人臉由於會有嚴重的扭曲,致使偵測率大幅下降。因此我們提出能改善一般人臉偵測器,使其能應用於偵測魚眼影像上的方法。我們的方法主要包含兩個部分:其一,藉由調整網路架構來區分出兩個學習目標,分別是中間區域人臉以及邊緣區域人臉的特徵學習;其二,我們提出根據人臉候選框所在的影像區域調整其形狀以及大小的方法。除了提高在魚眼影像邊緣區域的人臉偵測準確率以外,我們提出的人臉偵測器能直接應用於原始的魚眼影像上,而不須先將魚眼影像進行透視轉換或縫合校正。由於在魚眼相機的應用中,實時性非常重要,因此我們以一個能實時偵測的RFB人臉偵測器作為模型基底來實驗我們的方法。從偵測真實魚眼影像資料集的結果顯示,我們提出的方法能有效提高偵測器的平均精度(average precision)達5%,同時也能在CPU上維持與原模型相當的實時運行速度。 Tremendous progress has been made for face detection from normal images in recent years; however, accurate and fast face detection for fisheye images remains a challenging issue because of serious fisheye distortion in the peripheral region of the image. To improve face detection accuracy, we propose a light-weight face detector that uses a location-aware network to distinguish the peripheral region from the central region in the feature learning process. To match the face detector, the shape and scale of the anchor (bounding box) is made location dependent. The proposed method performs directly in the fisheye image domain without rectification and calibration; therefore, it is agnostic of the fisheye projection parameters. We test the proposed method on Wider-360 and real-world fisheye images using a single CPU core. The proposed method achieves superior accuracy over the state-of-the-art real-time face detector RFB Net at comparable speed. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/76612 |
DOI: | 10.6342/NTU202004192 |
Fulltext Rights: | 未授權 |
Appears in Collections: | 電信工程學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
U0001-3108202005525700.pdf Restricted Access | 2.28 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.