以快速收斂 UNET 深度學習之模型進行河床之影像分割

Performing Semantic Segmentation On Riverbed With FastConvergent UNET Deep Neural Network.

黃彥儒(Yen-Ju Huang)

指導教授 ： 韓仁毓

國立臺灣大學/工學院/土木工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201902238

摘要

在工程問題中，常常會利用影像分割的技術來幫助解決真實世界之工程問題，透過對於河床影像之分割來輔助獲取河道空間及屬性進而作為橋梁安全的評估即為本研究之重要工作之一。在傳統的影像處理技術中，要完成這種多群之分割，往往需要人工調整不同的分割演算法的參數;近年來，雖然深度學習的技術有了重大突破，於影像分割之應用也屢見不鮮，但是其模型的訓練上往往需要大量的時間、資料。本研究透過近年來深度學習幾個於電腦視覺領域中具突破性之訓練策略、模型設計以及超參數和超參數間之搭配來優化訓練演算法。此外，為了測試以及檢驗演算法之可用性以及性能，本研究利用2018年CVPR所辦比賽提供的空拍影像資料來做為本研究之比較、輔助驗證。透過本研究採用之訓練演算法，相較於深度學習傳統方式的訓練，不僅大幅減少至少三倍所需之訓練時間，其準確度更提高1~2%的精度。

關鍵字

深度學習 ； 影像分割 ； 快速收斂

並列摘要

In the real world problem, performing image semantic segmentation can be helpful to solve real world problem, for example, this technology could potentially apply on gaining information of riverbed as one of indicator for monitoring bridge state. In traditional image processing algorithm, we always need to tune the parameters in the algorithm manually. Deep learning technology has big breakthrough over the past few years, it’s often to see people apply it on image segmentation. However, it always takes lots of time and data for training the model. This research takes the combination of break-through training strategies of deep learning in computer vision field and seeking to decrease the time spending for training our model. Furthermore, we examine the training algorithm by two sets of data including ours and from the competition hold by CVPR in 2018. We successfully make our model converge three times faster than what it used to take and even outperform the model trained with traditional method by 1~2% of accuracy.

並列關鍵字

deep learning ； semantic segmentation ； fast convergence

參考文獻

Cortes, C. & Vapnik, V. (1995). Support Vector Networks. Machine Learning, 20, 273-297.

Dechter, R. (1986). Learning While Searching in Constraint- atisfaction- roblems. AAAI

Dinh, L., Pascanu, R., Bengio, S. & Bengio, Y. (2017). Sharp Minima Can Generalize For Deep Nets.. CoRR, abs/1703.04933.

Dauphin, Y. N., Pascanu, R., Gülçehre, Ç., Cho, K., Ganguli, S. & Bengio, Y. (2014). Identifying and attacking the saddle point problem in high-dimensional non-convex optimization.. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence & K. Q. Weinberger (eds.), NIPS (p./pp. 2933-2941), .

Duchi, J. C., Hazan, E. & Singer, Y. (2011). Adaptive Subgradient Methods for Online Learning and Stochastic Optimization.. Journal of Machine Learning Research, 12, 2121-2159.