多作物語義分割的集成模型：水稻和卷心菜｜國立中央大學博碩士論文系統

簡易檢索 / 詳目顯示

回結果列表

研究生：	蘇吉托 Adventus Ridwan Kis Sugiyarto
論文名稱：	多作物語義分割的集成模型：水稻和卷心菜 The Ensemble Model for Semantic Segmentation of Multi-crops: Rice and Cabbage
指導教授：	梁德容 De-Ron Liang
口試委員:
學位類別：	碩士 Master
系所名稱：	資訊電機學院 - 資訊工程學系 Department of Computer Science & Information Engineering
論文出版年：	2022
畢業學年度：	110
語文別：	英文
論文頁數：	59
中文關鍵詞：	語義分割、集成模型、UNet 、VGG16
外文關鍵詞：	semantic segmentation, ensemble model, UNet, VGG16
相關次數：	點閱：13 下載：0
分享至:	分享至facebook 分享至twitter

查詢本校圖書館目錄查詢臺灣博碩士論文知識加值系統勘誤回報

卷積神經網絡 (CNN) 可用於數字圖像處理，例如語義分割。使用此模型，可以分割航拍圖像以識別地塊中的作物。可以對航拍圖像執行語義分割的 CNN 架構之一是 UNet-VGG16。該架構用於 NCU RSS 1.3 架構中，該架構可以在航拍圖像中分割水稻，就像去年國立中央大學 (NCU) 之前的工作一樣。此外，目前有來自行政院農業委員會 (COA) 的另一種作物的捲心菜數據集，為 NCU 的下一次工作做好準備。

NCU RSS1.3 可用於訓練模型，該模型可以在航空圖像的包裹中預測水稻或捲心菜的多種作物。然而，集成方法也能夠預測航拍圖像中的作物，因為可以系統地添加新的作物數據集。集成模型將整合每種作物子模型的預測結果。如果新的作物數據集準備好，之前數據集的子模型可能不需要在集成模型中訓練。

本研究旨在比較一個集成模型基線，該基線將來自每個作物的一個類別的子模型與使用多類作物訓練的單一模型相結合。數據集分為五輪訓練、驗證和測試。 Cohen-Kappa 公式用於找出基本事實，並根據有關背景、大米和捲心菜的基於包裹的信息進行預測。第一個實驗是比較沒有調整減重函數的集成模型和多作物單模型。結果是未調整重量損失函數的集成模型的平均 kappa (0.917) 高於多作物單模型的平均 kappa (0.897)。第二個實驗是比較沒有調整權重損失函數和調整權重損失函數的集成模型。結果是調整權重損失函數的集成模型的平均 kappa (0.953) 高於沒有調整權重損失函數的值。與多作物單模型的 kappa 相比，具有調整權重損失函數的 ensemble 模型的 kappa 有顯著差異。此比較的 Wilcoxon 有符號秩 p 值為 0.035。結論是，在本研究中，具有調整權重損失函數的集成模型比多作物單模型具有更好的性能。這意味著集成模型改進了多作物單模型。

The Convolutional Neural Networks (CNN) can be used in digital image processing, such as semantic segmentation. Using this model, an aerial image can be segmented to identify the crop in the parcel field. One of the CNN architectures that can perform semantic segmentation for aerial images is UNet-VGG16. This architecture is used in the architecture NCU RSS 1.3 which can segment rice in aerial images as in previous work of National Central University (NCU) last year. Furthermore, currently there is a cabbage dataset for another crop from the Council of Agriculture (COA) Executive Yuan that is ready for next NCU’s work.

The NCU RSS1.3 can be used in training models that can predict multi-crops of rice or cabbage in parcels of aerial images. However, the ensemble approach is also able to predict the crop in the aerial image because new crop datasets may be systematically added. The ensemble model will integrate the prediction outcomes from each crop's sub-models. If the new crop dataset is ready, the sub-model of the previous dataset may not need to train in the ensemble model.

This study aims to compare an ensemble model baseline that combines sub-models from one class from each crop with the single model trained with multi-class crops. The dataset is split into train, validation, and test in five rounds. The Cohen-Kappa formula is used to figure out the ground truth and make predictions based on parcel-based information about the background, rice, and cabbage. The first experiment is comparing the ensemble model without tuning weight loss function and the multi-crops single model. The result is the average kappa of the ensemble model without tuning weight loss function (0.917) higher than the average kappa of the multi-crops single model (0.897). The second experiment is comparing the ensemble model without tuning weight loss function and with tuning weight loss function. The result is the average kappa of the ensemble model with tuning weight loss function (0.953) higher than without tuning weight loss function. The kappa of the ensemble model with tuning weight loss function is significantly different compared to the kappa of the multi-crops single model. The Wilcoxon Signed-Rank p-value of this comparison is 0.035. The conclusion is that the ensemble model with tuning weight loss function has better performance than the multi-crops single model in this research. It means that the ensemble model improves the multi-crops single model.

中文摘要        i
Abstract        ii
Acknowledgment    iii
Table of Contents    iv
List of Figures    vi
List of Tables        vii
Chapter I    Introduction    1
1.1.    Background    1
1.2.    Motivation and Objective    2
1.3.    Research Hypothesis    3
1.4.    Problem Statements    4
1.5.    Contribution    4
1.6.    Limitation of Study    4
1.7.    Thesis Structure    5
Chapter II    Literature Review    6
2.1.    Semantic Segmentation Using Unet-VGG16    6
2.2.    The Ensemble Model    7
Chapter III    Research Method    10
3.1.    Proposed Solution    10
3.1.1.    The Input    12
3.1.2.    Data Preprocessing    13
3.1.3.    The Training of Sub-Model    16
3.1.4.    Generate Segmentation Prediction    17
3.1.5.    Combine Sub-Models Using Combiner    18
3.1.6.    The Output    28
3.2.    Multi-Crops Single Model    30
Chapter IV    Experiments and Result Analysis    32
4.1.    Experiments    32
4.1.1.    The Dataset of Experiment    32
4.1.2.    The experiment of ensemble model compared to multi-crops single model    34
4.1.3.    The Experiment of Ensemble Model with Tuning Parameter    37
4.2.    Results & Analysis    37
4.2.1.    The Experiment Result of Ensemble Model Compared to Multi-crops Single Model    37
4.2.2.    The Experiment Result of Ensemble Model with Tuning Parameter    39
Chapter V    Conclusion    43
5.1.    Conclusion    43
5.2.    Future Works    44
Bibliography        45
                                

[1] Council of Agriculture, 2022. https://eng.coa.gov.tw/ws.php?id=9501.
[2] Y. Guo, Y. Liu, T. Georgiou, et al, "A review of semantic segmentation using deep neural networks," in Int J Multimed Info Retr 7, pp. 87-93, June 2018. https://doi.org/10.1007/s13735-017-0141-z.
[3] I. Wahyuni, W. -J. Wang, D. Liang, et al, "Rice Semantic Segmentation Using Unet-VGG16: A Case Study in Yunlin, Taiwan," 2021 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS), pp. 1-2, Dec. 2021, doi: 10.1109/ISPACS51563.2021.9651038.
[4] O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," in CoRR, vol. abs/1505.04597, May 2015, doi: 10.48550/arXiv.1505.04597.
[5] X. Yin, L. Sun, Y. Fu, R. Lu and Y. Zhang, "U-Net-Based Medical Image Segmentation," in J Healthc Eng, Apr. 2015, doi: 10.1155/2022/4189781.
[6] K.B. Soulami, N. Kaabouch, M.N. Saidi, et al, "Breast cancer: One-stage automated detection, segmentation, and classification of digital mammograms using UNet model based-semantic segmentation," in Biomedical Signal Processing and Control, vol. 66, Apr. 2021, doi: 10.1016/j.bspc.2021.102481.
[7] N. Enshaei, S. Ahmad and F. Naderkhani, "Automated detection of textured-surface defects using UNet-based semantic segmentation network," in 2020 IEEE International Conference on Prognostics and Health Management (ICPHM), 2020, pp. 1-5, doi: 10.1109/ICPHM49022.2020.9187023.
[8] Y. Sun, F. Bi, Y. Gao, et al, "A Multi-Attention UNet for Semantic Segmentation in Remote Sensing Images," in Symmetry, vol. 14, no. 5: 906, Apr. 2022, doi: 10.3390/sym14050906.
[9] K. Simonyan and A. Zisserman, "Very Deep Convolutional Networks for Large-Scale Image Recognition," in ICLR, Apr. 2015, doi: 10.48550/arXiv.1409.1556.
[10] S. Noppitak and O. Surinta, "Ensemble Convolutional Neural Network Architectures for Land Use Classification in Economic Crops Aerial Images," in ICIC Express Letters, vol. 15, pp. 531-543, June 2021, doi: 10.24507/icicel.15.06.531.
[11] D. Cao, H. Xing, M. S. Wong, et al, "A Stacking Ensemble Deep Learning Model for Building Extraction from Remote Sensing Images," in Remote Sensing, vol. 13(19), Sept. 2021, doi: 10.3390/rs13193898.
[12] A. Abdollahi, B. Pradhan, and A. Alamri, "An Ensemble Architecture of Deep Convolutional Segnet and Unet Networks for Building Semantic Segmentation from High-resolution Aerial Images," in Geocarto International, Dec 2020, doi: 10.1080/10106049.2020.1856199.
[13] J. Wang, and G. Song, "A Deep Spatial-Temporal Ensemble Model for Air Quality Prediction," in Neurocomputing, vol. 314, pp. 198-206, Nov. 2018, doi: 10.1016/j.neucom.2018.06.049.

簡易檢索 / 詳目顯示

相關論文