臺灣博碩士論文加值系統

人工智慧與深度學習現今已被廣泛使用在各個領域，無論是影像辨識、電
子商業，亦或是新興媒體產業 等， 其效能也遠遠高於以往之傳統處理方式。現
今網路世界，各個餐廳往往會將餐廳 照片放置網路，以供大眾參考，或是網美
們會將去過的餐廳分享至 Instagram 、 Facebook 等社群軟體 ，但照片內容琳瑯
滿目，沒有重點。所以我們想設計一個應用介面，能將雜亂的圖片做適當的分
類，讓使用者一目了然。
本論文使用卷積神經網路 VGG16 及 Inception ，配合 PCA Principal
components analysis 的降維，將雜亂的圖片簡單分為五類 食物、菜單、室內
場景、室外場景、其他，以 I magenet 作為 預訓練模型 ，再將從 G oogle map 抓
取之 90 000 張各式餐廳場景圖片當作 訓練資料 ，擷取之 全連接 6 層之特徵經過
PCA 降 維 處理後 輸入進 SVM support vector machine 做分類，得到最後之分類
結果。

Artificial intelligence and deep learning are now widely used in various fields, For example,image recognition, electronic commerce, or the new media industry... its performance is much higher than the traditional way. In today's online world, restaurants often place photos of restaurants on the Internet for public reference, or Internet celebrities will share the restaurants they have visited to social media such as Instagram and Facebook, but the photos are dazzling and unfocused. So we want to design an application that can properly classify cluttered images for the user to see at a glance.
This paper uses convolutional neural network-Inception, combined with PCA's dimensionality reduction, to divide the messy pictures into five categories: food, menu, indoor scene, outdoor scene, and others. Imagenet is used as the Pretrain model, and then from Google map. The 93,000 pictures of various restaurant scenes were used as training data. The features of the Global Pooling layer captured by the PCA were inputted to the SVM for classification and the final classification result was obtained.

目錄
摘要
i
英文摘要
ii
目錄
iii
圖目錄
iv
表目錄
v
第
1 章 緒論 1
1.1
研究動機 1
1.2
相關文獻 1
1.3
系統架構介紹 2
1.4
論文架構 2
第
2 章 背景知識與方法研究 3
2.1
Slim 模組 3
2.2
VGG16 卷積網路模型 3
2.3 Inception V4 卷積網路模型 5
2.4
PCA 主成分分析 8
2.5
SVM 分類器 9
第
3 章 研究方法與系統程式 10
3.1
場景資料庫取得 10
3.2
場景訓練資料之標籤 11
3.3
場景特徵取得與處理 11
3.3.1
資料預處理 11
3.3.2
建立網路模型 12
3.3.3
特徵提取 12
3.3.4
PCA 降維 13
3.3.5
SVM 分類 14
3.4
系統介面功能說明 15
第
4 章 實驗結果 18
4.1
餐廳影像資料庫 18
4.2
實驗數據及分析 18
4.2.1
實驗一 以 Inception 做特徵擷取網路之分類 18
4.2.2
實驗二 : 以 VGG16 做為特徵擷取網路之分類 21
4.2.3
實驗三： 以 Inception 做 End to end 網路分類 23
4.2.4
實驗四 : 以 Inception 做特徵擷取網路之不同場景分類 25
4.2.5
實驗結果 27
第
5 章 結論與未來研究方向 30

[1] Ariadna Quattoni,Antonio Torralba. “Recognizing indoor scenes.” IEEE Conference on Computer Vision and Pattern Recognition,2009
[2] Bolei Zhou1, Agata Lapedriza1, Jianxiong Xiao,Antonio Torralba1, Aude
Oliva1. “Learning Deep Features for Scene Recognition using Places
Database.” NIPS'14 Proceedings of the 27th International Conference on Neural
Information Processing Systems - Volume 1 Pages 487-495,2014
[3] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,
Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich.
“Going Deeper with Convolutions.” IEEE Conference on Computer Vision
and Pattern Recognition (CVPR),2015
[4] Karen Simonyan, Andrew Zisserman. “Very Deep Convolutional Networks for
Large-Scale Image Recognition.” arXiv preprint arXiv:1409.1556 ,2014. [5] Pearson, K.,”On Lines and Planes of Closest Fit to Systems of Points in Space.”, Philosophical Magazine. 2 (6): 559–572, 1901.
[6] Cortes, Corinna, and Vladimir VAPNIK. ”Support-vector networks. ”
Machine Learning, 20(3), 273–297, 1995
[7] Kaiming He,Georgia Gkioxari,Piotr Dollár,Ross Girshick.,”Mask R-CNN .”,
IEEE International Conference on Computer Vision (ICCV) , 2017.