隨著老年人口快速地增長，用藥量也隨之增加，伴隨而來的問題是病患因無法正 確地辨識外觀相似或名稱聽起來相似的藥物(look alike sound alike; LASA)進而造成病患服藥錯誤(patient medication errors, PEs)。為了解決此問題，近年來針對自動藥物辨 識的研究之重要性日漸提升。然而，對於形狀和大小相似的藥物容易辨識錯誤的問題在目前研究中仍未有效地被解決。因此，本論文利用機械手臂及旋轉盤擷取多視角拍攝的藥物影像，藉由藥物在不同視角可能呈現出獨一無二的外觀特徵，並結合卷積神經網路(convolutional neural network, CNN)自動萃取出有高鑑別力的藥物特徵，試圖解決在特定的單一視角下因藥物外觀相似而不易辨識的問題。
本論文主要分為藥物影像擷取、影像前處理和深度學習模型訓練。在影像擷取部 分，機械手臂擷取的藥物影像庫來自80種的藥錠，每種藥有64張影像，包含從藥物的正面、反面、斜側面和正側面的四種視角進行拍攝，此資料庫共5,120張影像;旋轉盤擷取的藥物影像庫來自100種的藥錠，每種藥物有160張影像，包含從藥物的正面和反面在旋轉時的不同視角進行拍攝，此資料庫共16,000張影像。在影像前處理部分，原始影像經由大津二值化法(Otsu’s thresholding)定位藥物區域進行以藥物為中心 的剪裁(image cropping)，再進一步將四張來自不同視角的藥物影像以四宮格的排列方式組合成一張多視角的影像。然而，由於組合後張數減少，因此透過四張影像在四宮格中不同位置的重組排列方式達到影像資料擴增(image augmentation)的目的。當影像擴增後，來自機械手臂的多視角藥物影像庫，每種藥可增至384張影像，整體共30,720張影像;來自轉盤的多視角藥物影像庫，每種藥可增至960張影像，整體共96,000 張影像。在模型訓練部分，本論文採用 AlexNet CNN預訓練的模型，透過遷移式學習(transfer learning)的方式來訓練模型，並以 4 折交叉驗證(4-fold cross validation)的方式來評估模型的穩定性和表現，最後以Top-k準確率呈現結果。
研究結果顯示，來自機械手臂和旋轉盤擷取的影像庫，以未組合的單一視角影像 進行模型訓練的表現，其Top-1準確率分別為99.39%和99.83%，經組合後的多視角影像庫，訓練後的模型表現，皆可提升至100%。因此，本研究提出的多視角影像結 合的方法搭配Alexnet CNN的預訓練模型進行自動藥物辨識的應用，可以有效地解決在特定單一視角下因藥物外觀特徵相似而不易辨識的問題。

With the rapid rise of the elderly population, the use of prescription pills has increased along with patient associated medication errors (PEs) resulting from patients who are unable to correctly recognize Look Alike Sound Alike (LASA) medications. To solve this problem, the development of an application for automatic pill recognition is essential. However, the problem of misclassification for pills with a similar appearance in terms of shape and color remained unsolved. Therefore, this thesis attempts to build the multi-perspective pill image database that may reveal unique features from different perspectives of the pill and utilize it to develop an automatic pill recognition based on convolutional neural network (CNN) in order to address the challenge for identifying pills with similar appearance under specific single perspective.
The proposed method is comprised of three parts including pill image acquisition, image preprocessing, and automatic pill recognition algorithm. For image acquisition, two independent image databases were acquired by leveraging the robotic arm and turntable plate. In the robotic arm acquired database, there are 80 types of oral tablets included. For each pill, it was captured under the perspective of the lateral side, tilted lateral side, front side, and back side. The number of total images in this database is 5,120. In the turntable plate acquired database, there are 100 types of oral tablets included. Each pill was captured from different perspectives of rotation angles. The number of total images in this database is 16,000. For image preprocessing, the pill area in the raw image was positioned to the center of the image using Otsu’s thresholding and image cropping was performed. To construct multi-perspective pill image data, four images from different perspectives were combined side by side to form a 2 × 2 tiled image. However, the number of total images reduced after combination, so the original four images within a tiled image underwent re-arrangement to different positions for image augmentation. After image augmentation, the number of images from the robotic arm acquired database increased to 384 images per pill and 30,720 images in total. For the turntable plate acquired database, it increased to 960 images per pill and 96,000 images in total. Regarding model training, the pre-trained AlexNet model was applied for automatic pill classification through transfer learning. The performance of the model was evaluated by 4-fold cross validation in terms of Top-k accuracy.
The experimental results showed the Top-1 accuracy of 99.39% and 99.83% achieved from the robotic arm acquired database and turntable plate acquired database, respectively, using single perspective of pill images to train the model. The classification accuracy can further improve to 100% for both databases using multi- perspective pill images to train the model. It shows that the proposed method utilizing multi-perspective pill images for CNN-based automatic pill classification can effectively address misclassification due to the similar appearance of the pills from a specific single perspective.

[1] World Health Organization, The third WHO Global Patient Safety Challenge: Medication Without Harm, March 29, 2017. Accessed on: Jun. 20, 2020. [Online]. Available: https://www.who.int/patientsafety/medication-safety/en/.
[2] A. Agrawal, “Medication errors: prevention using information technology systems,” British Journal of Clinical Pharmacology, vol. 67, no. 6, pp. 681–686, Jun. 2009. Accessed on: Jun. 20, 2020. [Online]. Available: http://doi.wiley.com/10.1111/j.1365-2125.2009.03427.x
[3] J. J.Mira, S.Lorenzo, M.Guilabert, I.Navarro, andV.Pérez-Jover, “A systematic review of patient medication error on self-administering medication at home,” Expert Opin. Drug Saf., vol. 14, no. 6, pp. 815–838, Jun.2015. Accessed on: Jun. 20, 2020. [Online]. Available: http://www.tandfonline.com/doi/full/10.1517/14740338.2015.1026326
[4] 顏兆熊、謝美芬, “老人多重用藥 - 台灣家庭醫學醫學會,” 家庭醫學與基層醫 療, vol. 24, no. 10, pp. 351–355, Oct. 25, 2009. Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.6965/FMPMC.200910.0351
[5] C. Bujang, W. Senior, M. Senior et al., Guide on handling look alike, sound alike medications, 1st ed. Pharmaceutical Services Division Ministry of Health Malaysia, 2012. Accessed on: Jun. 20, 2020. [Online]. Available: www.pharmacy.gov.my.
[6] National Institutes of Health, DailyMed, National Institutes of Health, National Library Medicine, U.S., Accessed on: Jun. 20, 2020. [Online]. Available: https://dailymed.nlm.nih.gov/dailymed/.
[7] National Institutes of Health, PillBox, National Institutes of Health, National Library Medicine, U.S., Accessed on: Jun. 20, 2020. [Online]. Available: https://pillbox.nlm.nih.gov/.
[8] WebMD, Pill Identifier - Find Pills by Color, Shape, Imprint, or Picture, Accessed on: Jun. 20, 2020. [Online]. Available: https://www.webmd.com/pill- identification/default.htm
[9] C. H. Hsieh, “Applying content cased image retrieval to drug identification research,” 2005. Accessed on: Jun. 20, 2020. [Online]. Available: https://hdl.handle.net/11296/576hs8
[10] “The GNU Image-Finding Tool.” Accessed on: Jun. 20, 2020. [Online]. Available: https://www.gnu.org/software/gift/.
[11] V. Tyagi, “Content-based image retrieval: an introduction,” in Content-Based Image Retrieval, Singapore: Springer, 2017, ch. 1, pp. 1–27, Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.1007/978-981-10-6759-4_1
[12] C. H. Lin, E. W. Huang, and W. W. Jiang, “Pill image retrieval using neural networks,” J. Taiwan Assoc. Med. Informatics, vol. 16, no. 2, pp. 29–42, 2007. Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.29829/TJTAMI.200706.0003.
[13] X. Jiang and A.Wah, “Constructing and training feed- forward neutral networks for pattern classification,” Pattern Recognit., vol. 36, no. 4, pp. 853–867, Apr. 2003. Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.1016/S0031- 3203(02)00087-0
[14] M. Kuei, “Visual pattern recognition by moment invariants.” IRE Transactions on Information Theory, vol. 8, no. 2, pp. 179-187, Feb. 1962. Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.1109/TIT.1962.1057692
[15] W. Kim and Y. Kim, “Region-based shape descriptor using Zernike moments,” Signal Process. Image Commun., vol. 16, no. 1, pp. 95–102, Sep.2000. Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.1016/S0923-5965(00)00019-9
[16] Y. H. Chen, “The study of automatic drug images identification system based on multiple images feature and adjust feature weights,” 2010. Accessed on: Jun. 20, 2020. [Online]. Available doi: https://hdl.handle.net/11296/4bu39c
[17] J. J. Caban, A. Rosebrock, and T. S. Yoo, “Automatic identification of prescription drugs using shape distribution models,” in 2012 19th IEEE International Conference on Image Processing, pp. 1005–1008, Sep., 2012, Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.1109/ICIP.2012.6467032.
[18] R. Osada, T. Funkhouser, B.Chazelle, and D. Dobkin, “Shape distributions.” ACM Transactions on Graphics, vol.21, no. 4, Oct., 2002, Accessed on: Jun. 20, 2020. [Online]. Available: https://doi.org/10.1145/571647.571648
[19] National Institutes of Health, Pillbox Rapid Identification, Reliable Information, National Institutes of Health, National Library Medicine, U.S., Accessed on: Jun. 20, 2020. [Online]. Available: https://pillbox.nlm.nih.gov/developer.html#data.
[20] Z. Chen and S. Kamata, “A new accurate pill recognition system using imprint information,” in Sixth International Conference on Machine Vision (ICMV 2013), vol. 9067, Nov. 2013, Accessed on: Jun. 20, 2020. [Online]. Available doi: 10.1117/12.2051168.
[21] B. Epshtein, E. Ofek, and Y. Wexler, “Detecting text in natural scenes with stroke width transform,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 2963–2970, 2010, Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/CVPR.2010.5540041
[22] S. Belongie, J. Malik, and J. Puzicha, “Shape matching and object recognition using shape contexts,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 4, pp. 509– 522, Apr. 2002., Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/34.993558
[23] J. Yu, Z. Chen, and S. I. Kamata, “Pill recognition using imprint information by two-step sampling distance sets,” in Proceedings - International Conference on Pattern Recognition, pp. 3156–3161, Dec. 2014, Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/ICPR.2014.544.
[24] C. Grigorescu and N. Petkov, “Distance sets for shape filters and shape recognition,” IEEE Trans. Image Process., vol. 12, no. 10, pp. 1274–1286, Oct. 2003., Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/TIP.2003.816010
[25] Z. Chen, J. Yu, S. Kamata, and J. Yang, “Accurate system for automatic pill recognition using imprint information,” IET Image Process., vol. 9, no. 12, pp. 1039–1047, Dec. 2015, Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1049/iet-ipr.2014.1007.
[26] P. Felzenszwalb and D. Huttenlocher, “Efficient belief propagation for early vision.”
Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, July, 19, 2004, Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/CVPR.2004.1315041
[27] P. Chupawa and K. Kanjanawanishkul, “Pill identification with imprints using a neural network,” Int. J. Eng. Technol., vol. 1, no. 1, pp. 30–35, 2015. Accessed on: Jun. 21, 2020. [Online]. Available: https://rb.gy/vrhtah
[28] M. Vieira Neto, J. de Souza, P. Reboucas Filho, and A. W. de O. Rodrigues, “CoforDes: an invariant feature extractor for the drug pill identification,” in 2018
IEEE 31st International Symposium on Computer-Based Medical Systems (CBMS), vol. 1, pp. 30–35. Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/CBMS.2018.00013
[29] T. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE Trans. Inf. Theory, vol. 13, no. 1, pp. 21–27, 1967. Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/TIT.1967.1053964
[30] V. Jakkula, “Tutorial on support vector machine (SVM),” Sch. of EECS, Wash. State Univ., vol. 37, 2006. Accessed on: Jun. 21, 2020. [Online]. Available: https://course.ccs.neu.edu/cs5100f11/resources/jakkula.pdf
[31] I. Rish, “An empirical study of the naïve bayes classifier,” 2001. Accessed on: Jun. 21, 2020. [Online]. Available: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.330.2788&rep=rep1&typ e=pdf
[32] National Institutes of Health, Pill Image Recognition Challenge: Submission Instructions, National Institutes of Health, National Library Medicine, U.S., Jan. 19, 2016. Accessed on: Jun. 21, 2020. [Online]. Available: https://pir.nlm.nih.gov/challenge/submission.html
[33] L. Cordeiro, J. Lima, A. Rocha Ribeiro, F. Bezerra, P. Reboucas Filho, and A. Rocha Neto, “Pill image classification using machine learning,” in 2019 8th Brazilian Conference on Intelligent Systems (BRACIS), pp. 556–561, 2019. Accessed on: Jun. 21, 2020. [Online]. Available doi: 10.1109/BRACIS.2019.00103
[34] Haykin and Simon, Neural Networks: A Comprehensive Foundation, 2nd ed. U.S.: Pearson Education, Junly 1998, ch. 4, pp. 178–274, Accessed on: Jun. 22, 2020. [Online]. Available: http://index-of.co.uk/Artificial- Intelligence/Neural%20Networks%20-
%20A%20Comprehensive%20Foundation%20-%20Simon%20Haykin.pdf
[35] X. Zeng, K. Cao, and M. Zhang, “MobileDeepPill: a small-footprint mobile deep
learning system for recognizing unconstrained pill images,” in Proceedings of the 15th Annual International Conference on Mobile Systems, Applications, and Services, pp. 56–67, 2017. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1145/3081333.3081336
[36] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep cnvolutional nural ntworks.” vol. 25, no. 2, Jan. 2012. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1145/3065386
[37] Y. Wong, H. Ng, K. Leung, K. Chan, S. Chan, and C. Loy, “Development of fine- grained pill identification algorithm using deep convolutional network,” J. Biomed. Inform., vol. 74, pp. 130–136, Oct.2017. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1016/j.jbi.2017.09.005
[38] T. Ho, “Random decision forests,” in Proceedings of the International Conference on Document Analysis and Recognition, vol. 1, pp. 278–282, 1995. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1109/ICDAR.1995.598994
[39] Y. Wang, J. Ribera, C. Liu, S. Yarlagadda, and F. Zhu, “Pill recognition using minimal labeled data,” in 2017 IEEE Third International Conference on Multimedia Big Data (BigMM), pp. 346–353, 2017. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1109/BigMM.2017.61
[40] C. Szegedy et al., “Going deeper with convolutions,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2015, pp. 1–9. Oct. 2015. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1109/CVPR.2015.7298594
[41] N. Larios Delgado et al., “Fast and accurate medication identification,” npj Digit. Med., vol. 2, no. 1, p. 10, Dec. 2019. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1038/s41746-019-0086-0
[42] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. pp. 770–778. Dec. 2016. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1109/CVPR.2016.90
[43] A. Howard et al., “MobileNets: efficient convolutional neural networks for mobile vision applications,” Apr. 2017. Accessed on: Jun. 22, 2020. [Online]. Available: http://arxiv.org/abs/1704.04861
[44] F. Iandola, S. Han, M. Moskewicz, K. Ashraf, W. Dally, and K. Keutzer, “SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,” Feb. 2016. Accessed on: Jun. 22, 2020. [Online]. Available: http://arxiv.org/abs/1602.07360
[45] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception architecture for computer vision,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 2818–2826, Dec. 2016. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1109/CVPR.2016.308
[46] W. Chang, L. Chen, C. Hsu, C. Lin, and T. Yang, “A deep learning-based Intelligent medicine recognition system for chronic patients,” IEEE Access, vol. 7, pp. 44441– 44458, 2019. Accessed on: Jun. 22, 2020. [Online]. Available doi: 10.1109/ACCESS.2019.2908843
[47] Y. Ou, A. Tsai, X. Zhou, and J. Wang, “Automatic drug pills detection based on enhanced feature pyramid network and convolution neural networks,” IET Comput. Vis., vol. 14, no. 1, pp. 9–17, Feb. 2020. Accessed on: Jun. 22, 2020. [Online].
Available doi: 10.1049/iet-cvi.2019.0171
[48] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi, “Inception-v4, inception-ResNet
and the impact of residual connections on learning,” in 31st AAAI Conference on Artificial Intelligence, pp. 4278–4284. 2017. Accessed on: Jun. 22, 2020. [Online]. Available: https://arxiv.org/abs/1602.07261v2
[49] W. Chang, L. Chen, C. Hsu, J. Chen, T. Yang, and C. Lin, “MedGlasses: a wearable smart-glasses-based drug pill recognition system using deep learning for visually impaired chronic patients,” IEEE Access, vol. 8, pp. 17013–17024, 2020. Accessed on: Jun. 22, 2020. [Online]. Available: https://ieeexplore.ieee.org/document/8962044/
[50] C. Xu and C. Xu, “A survey on multi-view learning,” 2013. Accessed on: Jun. 22, 2020. [Online]. Available: https://arxiv.org/pdf/1304.5634.pdf
[51] H. B. Mitchell, “Image Similarity Measures,” in Image Fusion: Theories, Techniques and Applications, Berlin, Heidelberg: Springer Berlin Heidelberg, 2010, pp. 167–185. Accessed on: Aug. 25, 2020 [Online]. Available: https://link.springer.com/chapter/10.1007/978-3-642-11216-4_14
[52] Sunny K. Tuladhar, K-means and PCA for Image Clustering: a Visual Analysis, Medium, June 3, 2020. Accessed on: Aug. 25, 2020 [Online]. Available: https://towardsdatascience.com/k-means-and-pca-for-image-clustering-a-visual- analysis-8e10d4abba40.
[53] Mathworks Contrast-limited adaptive histogram equalization (CLAHE) – adapthisteq, MATLAB, MathWorks, Accessed on: Jun. 01, 2020. [Online]. Available: https://www.mathworks.com/help/ [54] J. Schneider , “Cross validation.” Accessed on: Jun. 01, 2020. [Online]. Available: https://www.cs.cmu.edu/~schneide/tut5/node42.html.
[55] N. Otsu, “Threshold selection method from gray-level hostgrams.,” IEEE Trans Syst Man Cybern, vol. SMC-9, no. 1, pp. 62–66, 1979. Accessed on: Jun. 01, 2020. [Online]. Available doi: 10.1109/TSMC.1979.4310076.
[56] L. Taylor and G. Nitschke, “Improving deep learning using generic data augmentation,” Aug. 2017. Accessed on: Jun. 01, 2020. [Online]. Available: http://arxiv.org/abs/1708.06020.
[57] L. Perez and J. Wang, “The effectiveness of data augmentation in image classification using deep learning,” Dec. 2017. Accessed on: Jun. 01, 2020. [Online]. Available: http://arxiv.org/abs/1712.04621.
[58] I. G. and Y. B. and A. Courville, Deep Learning. MIT Press, 2016, ch. 6, pp. 180– 184, 189–191, ch. 9, pp. 327–329. Accessed on: Jun. 01, 2020. [Online]. Available: https://www.deeplearningbook.org/
[59] Mathworks, Visualize Activations of a Convolutional Neural Network, MATLAB, MathWorks, Accessed on: Jun. 10, 2020. [Online]. Available: https://www.mathworks.com/help/deeplearning/ug/visualize-activations-of-a- convolutional-neural-network.html
[60] H. Wu and X. Gu, “Towards dropout training for convolutional neural networks,” Neural Networks, vol. 71, pp. 1–10, Dec. 2015. Accessed on: Jun. 10, 2020. [Online]. Available doi: 10.1016/j.neunet.2015.07.007.
[61] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10. pp. 1345–1359, 2010. Accessed on: Jun. 10, 2020. [Online]. Available doi: 10.1109/TKDE.2009.191./ref/adapthisteq.html.
[62] D. Hemanth, B. Kumar, G. Manavalan, “Applications of memetic algorithms in image processing using deep learning,”in Recent Advances on Memetic Algorithms and its Applications in Image Processing, 1st ed. Singapore: Springer. 2020, ch. 4, pp. 69-91. Accessed on: Jun. 10, 2020. [Online]. Available doi: 10.1007/978-981- 15-1362-6.
[63] P. Kubben, M. Dumontier, and A. Dekker, Fundamentals of Clinical Data Science. Cham : Springer , 2019. Accessed on: Jun. 10, 2020. [Online]. Available: https://doi.org/10.1007/978-3-319-99713-1
[64] National Center for High-performance Computing (NCHC), Taiwan Computing Cloud (TWCC), Accessed on: Jun. 10, 2020. [Online]. Available: https://www.twcc.ai/
[65] Ru. Nagda, “Evaluating models using the Top N accuracy metrics,” Medium, 2018. Accessed on: Jun. 10, 2020. [Online]. Available: https://medium.com/nanonets/evaluating-models-using-the-top-n-accuracy-metrics- c0355b36f91
[66] C. Caragea et al., “Mean reciprocal rank (MRR),” in Encyclopedia of Database Systems, Boston: Springer, 2019. Accessed on: Jun. 10, 2020. [Online]. doi: 10.1007/978-0-387-39940-9_488
[67] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: unified, real-time object detection,” 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 779-788, 2016. Accessed on: Jun. 22, 2020. [Online]. Available: http://pjreddie.com/yolo/.
[68] J. Yan, Z. Lei, L. Wen, and S. Z. Li, “The fastest deformable part model for object detection,” 2014 IEEE Conference on Computer Vision and Pattern Recognition, pp. 2497-2504, 2014. Accessed on: Jun. 22, 2020. [Online]. doi:
10.1109/CVPR.2014.320
[69] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: towards real-time object
detection with region proposal networks,” arXiv, vol. 39, no. 6, Accessed on: Jan. 13, 2020. [Online]. Available: https://github.com/.