使用智慧型手機近距離自拍已是現今相當頻繁的行為，但在軟體與硬體飛速進步的當下，讓我們在拍照時可能已暴露在高度風險下而不自知。雖然人類的瞳孔在照片中所佔的面積不大，但若經過亮度、對比、飽和度調校以及影像超解析後，瞳孔反射影像所透露的資訊仍可能十分明顯，不僅影像中的物體經處理後更加清晰，透過人臉超解析的技術更能嘗試將低解析度的面部資訊還原，大大增加了曝露隱私的機率。由於人們的辨識能力相當強，僅藉由模糊的影像，只要對其夠熟悉，就可能識別出重要的訊息。本論文針對瞳孔反射影像進行超解析與相關處理，包括對影像進行物件辨識。針對距離、亮度與風險指數做出統計，並且還可針對有風險的照片加以處理並美化，使其拍照後的眼睛更明亮且不再有曝露風險的可能。

Taking selfies with a close distance using a smartphone is already most our daily activity. However, with the rapid advances on software and hardware, we have already been exposed with high risk when taking such pictures. Although a human pupil can only occupy a small proportion in a photo, after the adjustment of brightness, contrast, and saturation, together with image super-resolution, the information revealed by the pupil reflection could become obvious. Not only the object in the image could become clearer after the processing, but also the technique of facial super-resolution can partially restore the low-resolution facial information, thus greatly increasing the chance of exposing related privacy. As we human being are quite strong on recognizing things, it can be shown that only through a blurred image, as long as the involved person is relatively familiar, we could discover a great deal of important information. In this paper, super-resolution and related image processing of the pupil reflection image are implemented, and then the resulting image is used for object recognition to understand the potential risk. Statistics are collected based on distance, brightness and risk, so that we have a better understanding on how they co-related. To sum up, our system can assess the risk of an image with pupil reflection, replace the reflection image if necessary, and as a result, not only the corresponding eyes are made brighter, but also the potential risk of exposing privacy is completely removed.
keywords: pupil reflection, deep learning, super-resolution, object recognition, image security

[1]https://gendai.ismedia.jp/articles/-/68026
[2]https://news.cts.com.tw/cts/international/201910/201910091977403.html
[3]https://www.buzzfeed.com/jp/tatsunoritokushige/aragakiyuikawaii
[4]https://twitter.com/kose_sekkisei/status/753752677671067648
[5]Jenkins, R., & Kerr, C. (2013). Identifiable Images of Bystanders Extracted from Corneal Reflections
[6]Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018)
[7]Deokyun Kim, Minseon Kim, Gihyun Kwon, and Dae-Shik Kim. Progressive face super-resolution via attention to facial landmark. In Proceedings of the 30th British Machine Vision Conference ,BMVC(2019.)
[8]Faster R-CNN: Towards Real-Time ObjectDetection with Region Proposal Networks(2016)
[9]儀器儀表學報 2015年 第二期-安徽工程大學電氣工程學院 | 魏利勝 周聖文 張平改
[10]Konishino, Nayar, Corneal Imaging System Environment from Eyes(2006)
[11]Norman, Kent L./ Kirakowski, Jurek,The Wiley Handbook of Human Computer Interaction(2018)
[12]Andrew Yip, Pawan Sinha, Role of Color in Face
[13]Krizhevsky, Alex; Sutskever, llya; Hinton, Geoffrey E. “imagenet-classification-with-deep-convolutional-neural-networks”(2012)
[14]Christian Ledig, Lucas Theis, Ferenc Husz´ar, Jose Caballero, Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network(2017)
[15]https://ujjwalkarn.me/2016/08/11/intuitive-explanation-convnets/
[16]https://cs231n.github.io/convolutional-networks/
[17]https://github.com/jonbruner/generative-adversarial-networks/blob/master/gan-notebook.ipynb
[18]Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu,Chao Dong, Chen Change Loy, Yu Qiao, Xiaoou Tang, ESRGAN Enhanced Super-Resolution Generative Adversarial Networks (2018)
[19]https://github.com/idealo/image-super-resolution
[20]https://cloud.google.com/vision
[21]https://lens.google.com/
[22]https://www.digitimes.com.tw/tech/showimg.asp?sourcetype=2&filename=0000020190416_2__OTY49GXMO.png&source=20191025-234
[23]https://www.apple.com/tw/iphone/