本文使用MatConvNet及Caffe深度學習解決人工智慧圖像辨識並提供完整的解決方案，一個具備線上輸入的圖形辨識App在iOS裝置上。這個解決方案不僅將大型資料庫透過深度學習整合至卷積神經網路(CNN)，還將卷積神經網路跨計算平台的整合至iOS裝置上，完成圖像辨識App的應用程序設計。本文選擇可以在Matlab編譯環境運作的MatConvNet深度學習來建構與訓練CNN模型，相關作業環境的選擇有助於我們使用強大的數學工具與平行分散式處理分析大型的資料庫。目前iOS裝置皆已廣泛配備了語音功能、影像呈現與觸控螢幕等硬體設備，提供CNN圖像辨識App友善使用的測試環境。本文提供三個iOS App實例分別包含了，已發佈在蘋果應用軟體商店(App Store)的Handwriting 99 Multiplication、手寫英文字母的辨識及以分析BreakHis數據集為基礎的乳癌醫學圖像診斷。在安裝至iOS裝置前，每個圖像辨識所使用的CNN模型會先進行訓練與檢測。前兩個應用模型測試準確率分別為99.4%和97.0%，乳癌醫學影像診斷的兩個應用實例，分別是小葉癌(lobular carcinoma)與葉狀腫瘤(phyllodes tumor)的辨識、乳頭狀癌(papillary carcinoma)與乳腺腺病(adenosis)的辨識，經過數值實驗得到兩個例題的測試準確率分別為94.9%與87.3%。

This article presents a total solution to developing artificial intelligence and pattern recognition Apps on iOS devices using MatConvNet and Caffe deep learning. The solution integrates large scale data sets, deep learning and transformation of realized Convolutional Neural Networks (CNNs) across computational platforms toward App design on iOS devices. MatConvNet deep learning on Matlab programming environments facilitates constructing pattern recognition CNNs with powerful mathematical tools and parallel and distributed processes. The iOS devices provide pattern recognition CNN Apps friendly testing environments, which have been extensively equipped with modern audio, video, and screen-touching components. The iOS Apps presented here include the published handwriting 99 multiplication, handwritten English character classification, and medical image recognition of breast cancer derived from BreakHis datasets. The pattern recognition CNNs model of each App is tested before being mounted on iOS devices. The accurate rates for model testing of the first two Apps are respectively 99.4% and 97.0%, and diagnosing lobular carcinoma breast cancer against phyllodes tumor and papillary carcinoma against adenosis attains accuracy rate of 94.9% and 87.3% respectively.

第一章 Introduction 1
第一節 Pattern recognition CNN iOS Apps 1
第二節 Total Solution of integrating MatConvNet and Caffe deep learning for iOS App design 5

第二章 Architecture of Convolutional Neural Networks 7

第三章 Deep learning theory and Software 15
第一節 How does deep learning work? 15
第二節 Deep learning frameworks of MatConvNet and Caffe 24

第四章 Transformation across deep learning frameworks 27
第一節 iOS devices and Core ML 27
第二節 Caffe and Matcaffe 29
第三節 Learning by MatConvNet and executing on iOS devices 34

第五章 Handwriting 99 Multiplication on App Store 37
第一節 Dataset 37
第二節 CNN architecture 38
第三節 Training strategies 44
第四節 Numerical Experiments 44
第五節 Execution on iOS devices and publishing to App Store 45

第六章 Recognition of handwritten English characters 49
第一節 Dataset 49
第二節 CNN architecture 52
第三節 Training strategies 57
第四節 Numerical Experiments 59
第五節 Execution on iOS devices and publishing to App Store 59


第七章 Diagnosis of breast cancers by medical image recognition 65
第一節 Dataset 65
第二節 CNN architecture 69
第三節 Training strategies 73
第四節 Numerical Experiments 76
第五節 Execution on iOS devices and future App publication 77

第八章 Conclusions 79

[1]Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton,
“ImageNet classification with deep convolutional neural networks,”
In Proc. Advances in Neural Information Processing Systems 25., 2012, pp.1097-1105.

[2]Andrea Vedaldi, Karel Lenc,
“MatConvNet: Convolutional Neural Networks for MATLAB,”
Proceedings of the 23rd ACM international conference on Multimedia,
October 26-30, 2015, Brisbane, Australia [doi>10.1145/2733373.2807412]

[3]Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer,
“Automatic differentiation in PyTorch,”
In NIPS 2017 Autodiff Workshop: The Future of Gradient-based Machine Learning Software and Techniques, Long Beach, CA, US, December 9, 2017., 2017.

[4]Apple official,
“Converting Trained Models to Core ML,” [Online],
Available: https://developer.apple.com/documentation/coreml/converting\_trained\_models\_to\_core\_ml,
Accessed on: Jan. 07, 2018.

[5]C. Charalambous,
“Conjugate-gradient algorithm for efficient training of artificial neural networks,”
Inst. Electr. Eng. Proc.-G Circuits Devices Syst., vol. 139, pp. 301–310, 1992.

[6]Szegedy, C. et al.,
“Going deeper with convolutions,”
Preprint at http://arxiv.org/abs/1409.4842 (2014).

[7]Collin Hundley. (2017).,
“NeuralNet-Handwriting-iOS,” Swift AI on GitHub. [Online].
Available: https://github.com/Swift-AI/NeuralNet-Handwriting-iOS,
Accessed on: Jul. 31, 2017.

[8]Chao-Yuan, Tien. (2018).
“Handwriting 99 Multiplication,” iOS App Store. [Online].
Available: https://itunes.apple.com/us/app/handwriting-99-multiplication/id1419757476?mt=8,
Accessed on: Dec. 31, 2018.

[9]Francois Chollet and others. (2015).
“Keras,” GitHub. [Online].
Available: https://github.com/fchollet/keras,
Accessed on: Jan. 7, 2019.

[10]F. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte,
“A Dataset for Breast Cancer Histopathological Image Classification,”
IEEE Transactions of Biomedical Engineering, 2015.

[11]F. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte,
“Breast cancer histopathological image classification using convolutional neural networks,” In International Joint Conference on Neural Networks (2016).

[12]Cohen, G., Afshar, S., Tapson, J., & van Schaik, A. (2017).
“EMNIST: an extension of MNIST to handwritten letters,”
Retrieved from http://arxiv.org/abs/1702.05373

[13]Bridle, John S. (1990a). Soulié F.F.; Hérault J., eds.,
“Probabilistic Interpretation of Feedforward Classification Network Outputs, with Relationships to Statistical Pattern Recognition.,”
Neurocomputing: Algorithms, Architectures and Applications (1989).
NATO ASI Series (Series F: Computer and Systems Sciences). 68. Berlin, Heidelberg: Springer. pp. 227–236. doi:10.1007/978-3-642-76153-9_28.

[14]}Karen Simonyan, Andrew Zisserman,
“Very Deep Convolutional Networks for Large-Scale Image Recognition,”
arXiv:1409.1556 [cs.CV]
(Submitted on 4 Sep 2014 (v1), last revised 10 Apr 2015 (this version, v6))

[15]K. He, X. Zhang, S. Ren, and J. Sun,
“Deep residual learning for image recognition,”
In Proceedings of CVPR, pages 770–778, 2016.
arxiv.org/abs/1512.03385.

[16]M. T. Hagan and M. B. Menhaj,
“Training feedforward networks with the Marquardt algorithm,”
IEEE Trans. Neural Netw., vol. 5, no. 6, pp.989–993, Nov. 1994.

[17]M. Abadi, A. Agarwal et al.,
“Tensorflow: Large-scale machine learning on heterogeneous distributed systems,”
2016.

[18]Maruti Techlabs. (2018).
“8 Best Deep Learning Frameworks for Data Science enthusiasts,” [Online].
Available: https://medium.com/the-mission/8-best-deep-learning-frameworks-for-data-science-enthusiasts-d72714157761
Accessed on: Dec 12, 2018.

[19]Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986).,
“Learning internal representations by error propagation.,”
In D. E. Rumelhart, & J. L. McClelland (Eds.),
Parallel distributed processing, vol. 1 (pp. 318–362). MIT Press.

[20]R. Girshick, J. Donahue, T. Darrell, and J. Malik. %
“Rich feature hierarchies for accurate object detection and semantic segmentation,”
In CVPR, 2014.

[21]R. Girshick,
“Fast R-CNN,”
arXiv:1504.08083, 2015.

[22]Scherer, D., Müller, A., & Behnke, S.,
“Evaluation of pooling operations in convolutional architectures for object recognition.,”
In Proc. International conference on artificial neural networks (pp. 92–101), 2010.

[23]Sergey loffe, Christian Szegedy,
“Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,”
arXiv:1502.03167, (Submitted on 11 Feb 2015 (v1), last revised 2 Mar 2015 (this version, v3))

[24]Nair, Vinod and Hinton, Geoffrey E.,
“Rectified linear units improve restricted boltzmann machines.,”
In ICML, pp.807–814. Omnipress, 2010.

[25]W. Liu, Y. Wen, Z. Yu, and M. Yang.,
“Large-margin softmax loss for convolutional neural networks.,”
In ICML, 2016.

[26]World Health Organization(WHO).
“Breast cancer,” [Online].
Available: https://www.who.int/cancer/prevention/diagnosis-screening/breast-cancer/en/,
Accessed on: Dec. 01, 2018.

[27]Y. LeCun, K. Kavukcuoglu, and C. Farabet,
“Convolutional Networks and Applications in Vision,”
in Proceedings of 2010 IEEE International Symposium on Circuits and Systems
(ISCAS). Springer-Verlag, Jun. 2010, pp. 253–256.

[28]Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, Trevor Darrell,
“Caffe: Convolutional Architecture for Fast Feature Embedding,”
Proceedings of the 22nd ACM international conference on Multimedia, November 03-07, 2014, Orlando,
Florida, USA [doi>10.1145/2647868.2654889]

[29]LeCun, Y. (1988).,
“A theoretical framework for back-propagation.,”
In D. Touretzky, G. Hinton, \& T. Sejnowski (Eds.),
Proceedings of the 1988 connectionist models summer school (pp. 21–28). CMU, Pittsburgh, Pa: Morgan Kaufmann.

[30]Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard and L. D. Jackel,
“Backpropagation Applied to Handwritten Zip Code Recognition,”
Neural Computation, 1(4):541-551, Winter 1989.

[31]Y. Lecun, L. Bottou, Y. Bengio, P. Haffner
“Gradient-based learning applied to document recognition,”
Proceedings of the IEEE, ( Volume: 86 , Issue: 11 , Nov 1998 ) [doi> 10.1109/5.726791].

[32]Y. LeCun, Y. Bengio,
“Convolutional networks for images speech and time-series,”
in The Handbook of Brain Theory and Neural Networks, USA, MA, Cambridge: MIT Press, 1995.

[33]Yann LeCun, Yoshua Bengio and Geoffrey Hinton,
“Deep Learning,”
Nature, vol 521, 436-444, (28 May 2015).

[34]Kim, Yoon.,
“Convolutional neural networks for sentence classification,”
arXiv preprint ,arXiv:1408.5882.(2014)