白天在太陽光低角度直射時，無論是人類的眼睛還是攝影機鏡頭，會因強光直入產生逆光效果(Backlight Effect)而刺眼，甚或阻擋視線，逆光效果對於拍攝人景的方面來說或許是增添了些美感，但是對路上行車則會產生安全威脅。攝影機在自駕車上的應用非常多元，像用在辨識車道、燈號、及障礙物等，均對行車安全有直接的影響，而上述逆光現象會降低攝影機影像的辨識度而產生安全疑慮。本研究嘗試以深度學習模型將影像的逆光效果降到最低，利用自編碼器模型的特性，以消除雜訊為原理，將具逆光效果之影像轉換成正常的影像，試著將影像內被強光遮擋的部分還原回來，讓影像更清晰可見，如此影像在應用上即可將安全疑慮降到最低。
本研究採非監督式學習的方式，使用自編碼器(Auto-encoder)作為模型架構，訓練集為自製的資料集，為從多部行車紀錄器影片擷取出的正常無光害的影像，其中各影片的街景均為不同，但因為模型為非監督式學習，故不需要做任何的標註(label)。測試集亦是先從上述之影片中經過人工挑出有逆光效果出現的景象，另外再利用電腦修圖軟體Photoshop從正常影像裡挑選一些用來合成成具有不同程度逆光效果的影像，來補償測試樣本數的不足。訓練時，亦使用資料擴增的方式來增加訓練集的影像數量，彌補訓練集資料量不足的困擾。模型在學習大量正常無逆光效果的影像特徵後，將測試集丟入模型裡，即可將影像內具有的逆光特徵消除掉，以降低逆光效果，並還原為清晰影像以利後續辨識。研究亦將原始影像分別跟合成後的影像及模型生成後出的影像以結構相似性指標來做比對，以評估模型效能，模型所生成影像評分的平均數值大約落在0.9，分數高於測試集的影像，由此證明此模型的效果及可行性。

In the daytime, when the sun shines at a low angle may causes the backlight effect, the direct strong light will be dazzling or even block the sight of human eyes or camera lens. The backlight effect may be a little more aesthetically pleasing when taking photos, but it will pose a safety threat when driving on the road. Usages of cameras in self-driving cars are becoming essential today. For example, they are used to identify lanes, traffic lights, and other road obstacles, etc., and have a direct impact on driving safety. The backlight effect will reduce the resolution of camera images that will compromise the driving safety. This study attempts to use a deep learning model to minimize the backlight effect of images. Using the characteristics of the popular auto-encoder model, based on the principle of eliminating noise, tries to convert images with the backlight effect into normal images. The part of an image blocked by the backlight effect is restored to make the more clearly visible image that can be used for identifying objects and minimizing safety concerns.
This study adopts an unsupervised learning approach by using an autoencoder as the model architecture with a self-made training data set containing normal, light pollution-free images captured from a number of driving recorder videos found on the Internet. Data augmentation is used to increase the number of images in the training set to make up for the lack of images in the training set. The test set contains images with backlight effects manually selected from the same driving recorder videos and, to compensate for the insufficient number of test images, is augmented with images processed by computer retouching software by adding different degrees of backlight effects. After the model is trained with the normal images without backlight effect in the training dataset, it is tested by the test dataset to evaluate the degree of the backlight features elimination, i.e. to see how much the backlight effect is reduced to restore a clear image for subsequent objects recognition. This study also compared the original images with both the augmented processed images and the images generated from them by the model by using the structural similarity index to evaluate the performance of the model. The average score of the images generated by the model when compared to the origins is over 0.9, which is higher than when compared to the images in the test dataset. This confirms the performance and the feasibility of the model in this study.

[1] Yicheng Wu, Qiurui He, Tianfan Xue, Rahul Garg, Jiawen Chen, Ashok Veeraraghavan, Jonathan T. Barron. How to Train Neural Networks for Flare Removal. 2021 IEEE/CVF International Conference on Computer Vision (ICCV).
[2] Patricia Vitoria, Coloma Ballester. Automatic Flare Spot Artifact Detection and Removal in Photographs. arXiv:2103.04384v1 [eess.IV] 7 Mar 2021.
[3] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, Pierre-Antoine Manzagol. Extracting and Composing Robust Features with Denoising Autoencoders. Machine Learning, Proceedings of the Twenty-Fifth International Conference (ICML 2008), Helsinki, Finland, June 5-9, 2008.
[4] Rui Qian, Robby T. Tan, Wenhan Yang , Jiajun Su , and Jiaying Liu. Attentive Generative Adversarial Network for Raindrop Removal from A Single Image. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition.
[5] Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, Ying Wu. Semi-supervised Transfer Learning for Image Rain Removal. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 3877-3886.
[6] Hong Wang, Yichen Wu, Minghan Li, Qian Zhao, and Deyu Meng. A Survey on Rain Removal from Video and Single Image. arXiv:1909.08326v2 [eess.IV] 3 Oct 2019.
[7] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, John Paisley. Removing rain from single images via a deep detail network. July 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
[8] Xugang Lu, Yu Tsao, Shigeki Matsuda, Chiori Hori. Speech Enhancement Based on Deep Denoising Autoencoder. 2013 in INTERSPEECH (Conference of the International Speech Communication Association)
[9] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. MICCAI, 2015.
[10] Vincent Dumoulin, Francesco Visin. A guide to convolution arithmetic for deep learning. arXiv:1603.07285v2 [stat.ML] 11 Jan 2018.
[11] Ruijie Quan, Xin Yu, Yuanzhi Liang, Yi Yang. Removing Raindrops and Rain Streaks in One Go. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).
[12] Asha C S, Sooraj Kumar Bhat, Deepa Nayak, Chaithra Bhat. Auto Removal of Bright Spot from Images Captured Against Flashing Light Source. 2019 IEEE International Conference on Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER).
[13] GUANSONG PANG, CHUNHUA SHEN, LONGBING CAO, ANTON VAN DEN HENGEL. Deep Learning for Anomaly Detection: A Review. arXiv:2007.02500v3 [cs.LG] 5 Dec 2020.
[14] Adam Polyak, Lior Wolf, Yossi Adi, Yaniv Taigman. Unsupervised Cross-Domain Singing Voice Conversion. arXiv:2008.02830v1 [eess.AS] 6 Aug 2020.
[15] Retrieve from：https://towardsdatascience.com/autoencoders-and-the-denoising-feature-from-theory-to-practice-db7f7ad8fc78
[16] Retrieve from：https://zh.m.wikipedia.org/zh-tw/%E7%B5%90%E6%A7%8B%E7%9B%B8%E4%BC%BC%E6%80%A7
[17] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol. Stacked Denoising Autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion. Journal of Machine Learning Research 10 Dec 2010.