Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/98248
標題: 基於資料增量之多角度臉部辨識系統及其於即時影像追蹤之應用
Multi-Angle Face Recognition based on Data Augmentation and It's Application to Realtime Image Search
作者: 王慶裕
Ching-Yu Wang
關鍵字: 臉部識別
資料增量
Face Recognition
Data Augmentation
引用: [1] Z. L. S. M. I. Kaipeng Zhang, Zhanpeng Zhang and I. Yu Qiao, Senior Member,'Joint face detection and alignment using multi-task cascaded convolutional networks,' 2016. [2] J. P. Florian Schroff, Dmitry Kalenichenko, 'Facenet: A unified embedding for face recognition and clustering,' 2015. [3] X. S. Y. W. X. Z. Yao Feng, Fan Wu, 'Joint 3d face reconstruction and dense alignment with position map regression network,' 2018. [4] J. Hsu, 'Finding one face in a million.' [5] S. Z. Jiankang Deng, Jia Guo, 'Arcface: Additive angular margin loss for deep face recognition,' pp. 1–13, 2018. [6] Y. S. Q. D. G. S. Ren Wu, Shengen Yan, 'Deep image: Scaling up image recognition,' pp. 1–9, 2015. [7] T. V. Volker Blanz, 'A morphable model for the synthesis of 3d faces,' pp. 187–194, 1999. [8] S. Y. L. Z. H. Z. W. G. Dalong Jiangab, Yuxiao Hu, 'Efficient 3d reconstruction for face recognition,' pp. 787–798, 2004. [9] J. L. Q. Z. Feng Liu, Dan Zeng, 'Cascaded regressor based 3d face reconstruction from a single arbitrary view image,' pp. 1–8, 2015. [10] V. A. G. T. Aaron S. Jackson, Adrian Bulat, 'Large pose 3d face reconstruction from a single image via direct volumetric cnn regression,' 2017. [11] T. L. R. H. Rui Huang, Shu Zhang, 'Beyond face rotation: Global and local perception gan for photorealistic and identity preserving frontal view synthesis,' pp. 1–11, 2017. [12] 盧禮賓, '人在做監視器在看 全台16.7 萬支北市最密集.' [13] 胡瑋佳, '警政署下一步靠大數據分析破案,整合7 萬支監視器全臺大追緝.' [14] M. Y. M. R. L. W. Yaniv, Taigman, 'Deepface: Closing the gap to human-level performance in face verification,' pp. 1701–1708, 2014. [15] T. B. Z. W. C. H. Jingtuo Liu, Yafeng Deng, 'Targeting ultimate accuracy: Face recognition via deep embedding,' pp. 1–5, 2015. [16] Z. Y. M. L. B. R. L. S. Weiyang Liu, Yandong Wen, 'Sphereface: Deep hypersphere embedding for face recognition,' pp. 1–13, 2018. [17] Z. Z. X. J. D. G. J. Z. Z. L. W. L. Hao Wang, Yitong Wang, 'Cosface: Large margin cosine loss for deep face recognition,' pp. 1–11, 2018. [18] J.-S. H. X.-D. Z. X. Z. Jiang-Jing Lv, Xiao-Hu Shao, 'Data augmentation for face recognition,' pp. 184–196, 2017. [19] B.-N. C. J.-B. J. L. M. Daniel Crispell, Octavian, 'Dataset augmentation for pose and lighting invariant face recognition,' pp. 1–9, 2017. [20] Z. L. H. S. X. W. S. Z. L. Shifeng Zhang, Xiangyu Zhu, 'Faceboxes: A cpu real-time face detector with high accuracy,' pp. 1–9, 2017. [21] S. W. Hong-Wei Ng, 'A data-driven approach to cleaning large face datasets,' pp. 343–347, 2014. [22] W. X. O. M. P. A. Z. Qiong Cao, Li Shen, 'Vggface2: A dataset for recognising faces across pose and age,' 2018. [23] G. E. H. Alex Krizhevsky, Ilya Sutskever, 'Imagenet classification with deep convolutional neural networks,' pp. 1–9, 2012. [24] Y. S. Q. D. G. S. Ren Wu, Shengen Yan, 'Recognition of blurred faces using local phase quantization,' pp. 1–4, 2008. [25] J. D. Alejandro Newell, Kaiyu Yang, 'Stacked hourglass networks for human pose estimation,' 2016. [26] T. D. Jonathan LongEvan Shelhamer, 'Fully convolutional networks for semantic segmentation,' 2015. [27] S. R. J. S. Kaiming He, Xiangyu Zhang, 'Deep residual learning for image recognition,' 2015. [28] V. V. A. A. Christian Szegedy, Sergey Ioffe, 'Inception-v4, inception-resnet and the impact of residual connections on learning,' 2016.
摘要: 在本研究中,將模擬警方偵查案件情境,假設有一名嫌疑人,只有一張臉 部照片當作線索,希望透過一萬台監視器即時取得該秒一萬個影像辨識嫌疑犯並找出嫌疑人所在位置。文中將透過MTCNN [1] 擷取畫面中臉部特徵,並使用FaceNet [2] 識別嫌疑人。在缺乏嫌犯臉部照片情況下,人臉識別方法只能透過兩張照片轉成特徵向量計算相似度距離找出最相似的照片。在200 張的實驗下能有87% 的準確度判斷出嫌疑人。在2000 張的實驗下,準確度剩下79.2%。當照片提升至10000 張時,準確度則下降到45.1%。文中藉由PRNet [3] 實現2D 照片3D臉部建模,透過3D 模型增量產出各角度臉部照片,並訓練分類器,將臉部辨識準確度由45.1% 提升至71.4%。
In the study, the situation of the police investigation case will be simulated. Suppose there is a suspect with only one face photo as a clue. I hope to get 10,000 images to identify the suspect and find the suspect through 10,000 monitors. The face features in the picture will be captured through MTCNN [1] and the suspect will be identified using FaceNet [2]. In the absence of suspected face photos, the face recognition method can only find the most similar photos by calculating the similarity distance by converting two photos into feature vectors. The suspect can be judged with an accuracy of 87% in 200 experiments. Under the 2000 experiment, the accuracy is 79.2%. When the photo is increased to 10,000, the accuracy drops to 45.1%. In this paper, 2D photo 3D face modeling is implemented by PRNet [3], and the face images of each angle are incrementally generated through the 3D model, and the classifier is trained to increase the face recognition accuracy from 45.1% to 71.4%.
URI: http://hdl.handle.net/11455/98248
文章公開時間: 2021-11-15
Appears in Collections:資訊科學與工程學系所

文件中的檔案:

取得全文請前往華藝線上圖書館



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.