-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy paththesis.toc
More file actions
executable file
·99 lines (99 loc) · 8.35 KB
/
Copy paththesis.toc
File metadata and controls
executable file
·99 lines (99 loc) · 8.35 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
\babel@toc {italian}{}\relax
\babel@toc {italian}{}\relax
\contentsline {chapter}{Prefazione }{ii}{chapter*.1}%
\contentsline {chapter}{Indice}{iii}{chapter*.2}%
\contentsline {chapter}{\numberline {1}Introduzione}{1}{chapter.1}%
\contentsline {section}{\numberline {1.1}Motivazione }{1}{section.1.1}%
\contentsline {subsection}{\numberline {1.1.1}Modelli neurali per la segmentazione nel controllo qualità }{1}{subsection.1.1.1}%
\contentsline {subsection}{\numberline {1.1.2}Il dataset, requisiti e problematiche di realizzazione }{2}{subsection.1.1.2}%
\contentsline {subsection}{\numberline {1.1.3}Approccio al problema }{3}{subsection.1.1.3}%
\contentsline {subsubsection}{Generatore con architettura a solo decoder}{4}{section*.4}%
\contentsline {subsubsection}{Generatore con architettura a encoder-decoder}{5}{section*.6}%
\contentsline {section}{\numberline {1.2}La nascita del deep learning }{6}{section.1.2}%
\contentsline {subsection}{\numberline {1.2.1}Dal machine learning al deep learning}{6}{subsection.1.2.1}%
\contentsline {subsection}{\numberline {1.2.2}Le reti neurali biologiche}{7}{subsection.1.2.2}%
\contentsline {section}{\numberline {1.3}Le reti neurali feed forward}{7}{section.1.3}%
\contentsline {subsection}{\numberline {1.3.1}Il neurone artificiale}{8}{subsection.1.3.1}%
\contentsline {subsection}{\numberline {1.3.2}Il Back-propagation}{10}{subsection.1.3.2}%
\contentsline {subsection}{\numberline {1.3.3}Teorema di approssimazione universale}{12}{subsection.1.3.3}%
\contentsline {section}{\numberline {1.4}Le reti neurali convoluzionali}{13}{section.1.4}%
\contentsline {subsection}{\numberline {1.4.1}Storia delle CNN}{13}{subsection.1.4.1}%
\contentsline {subsection}{\numberline {1.4.2}La convoluzione}{15}{subsection.1.4.2}%
\contentsline {subsection}{\numberline {1.4.3}I parametri della convoluzione}{18}{subsection.1.4.3}%
\contentsline {subsubsection}{Padding}{18}{section*.20}%
\contentsline {subsubsection}{Stride}{18}{section*.21}%
\contentsline {subsubsection}{Kernel size}{19}{section*.22}%
\contentsline {subsubsection}{Numero di filtri}{19}{section*.24}%
\contentsline {subsubsection}{La dimensione di output}{20}{section*.25}%
\contentsline {subsection}{\numberline {1.4.4}Il pooling}{20}{subsection.1.4.4}%
\contentsline {subsubsection}{Max pooling}{21}{section*.26}%
\contentsline {subsubsection}{Average pooling}{21}{section*.28}%
\contentsline {subsection}{\numberline {1.4.5}La convoluzione multichannel}{22}{subsection.1.4.5}%
\contentsline {chapter}{\numberline {2}Stato dell'arte}{23}{chapter.2}%
\contentsline {section}{\numberline {2.1}Il primo modello GAN}{23}{section.2.1}%
\contentsline {subsection}{\numberline {2.1.1}L'adversarial training}{23}{subsection.2.1.1}%
\contentsline {subsection}{\numberline {2.1.2}La loss function}{25}{subsection.2.1.2}%
\contentsline {subsection}{\numberline {2.1.3}La convergenza del generatore}{27}{subsection.2.1.3}%
\contentsline {subsection}{\numberline {2.1.4}Il gradient vanishing}{28}{subsection.2.1.4}%
\contentsline {subsection}{\numberline {2.1.5}Il mode collapse}{28}{subsection.2.1.5}%
\contentsline {subsection}{\numberline {2.1.6}Alcuni risultati}{29}{subsection.2.1.6}%
\contentsline {section}{\numberline {2.2}DCGAN}{30}{section.2.2}%
\contentsline {subsection}{\numberline {2.2.1}L'architettura del modello}{30}{subsection.2.2.1}%
\contentsline {subsection}{\numberline {2.2.2}L'algebra vettoriale nello spazio latente Z}{31}{subsection.2.2.2}%
\contentsline {section}{\numberline {2.3}Wasserstein GAN}{32}{section.2.3}%
\contentsline {subsection}{\numberline {2.3.1}Earth-Mover distance}{33}{subsection.2.3.1}%
\contentsline {subsection}{\numberline {2.3.2}Alcuni risultati}{35}{subsection.2.3.2}%
\contentsline {section}{\numberline {2.4}Pix2Pix}{36}{section.2.4}%
\contentsline {subsection}{\numberline {2.4.1}La loss function}{36}{subsection.2.4.1}%
\contentsline {subsection}{\numberline {2.4.2}L'architettura}{37}{subsection.2.4.2}%
\contentsline {subsection}{\numberline {2.4.3}Alcuni esempi di applicazione}{38}{subsection.2.4.3}%
\contentsline {section}{\numberline {2.5}LaMa}{39}{section.2.5}%
\contentsline {subsection}{\numberline {2.5.1}Struttura del modello}{39}{subsection.2.5.1}%
\contentsline {subsubsection}{Fast Fourier Convolution}{40}{section*.52}%
\contentsline {subsection}{\numberline {2.5.2}La loss function}{41}{subsection.2.5.2}%
\contentsline {subsubsection}{La perceptual loss PL e la High receptive field perceptual loss HRFPL}{41}{section*.54}%
\contentsline {subsubsection}{L'adversarial loss}{42}{section*.56}%
\contentsline {subsubsection}{L'R1 regularization}{43}{section*.57}%
\contentsline {subsubsection}{La loss finale}{44}{section*.59}%
\contentsline {subsection}{\numberline {2.5.3}Alcuni risultati}{45}{subsection.2.5.3}%
\contentsline {chapter}{\numberline {3}Materiali e metodi }{46}{chapter.3}%
\contentsline {section}{\numberline {3.1}Il Dataset: Severstal steel defect detection }{46}{section.3.1}%
\contentsline {subsection}{\numberline {3.1.1}Distribuzione del dataset}{47}{subsection.3.1.1}%
\contentsline {section}{\numberline {3.2}Librerie e Framework }{51}{section.3.2}%
\contentsline {subsection}{\numberline {3.2.1}Numpy }{51}{subsection.3.2.1}%
\contentsline {subsection}{\numberline {3.2.2}OpenCV }{51}{subsection.3.2.2}%
\contentsline {subsection}{\numberline {3.2.3}Pytorch }{51}{subsection.3.2.3}%
\contentsline {subsection}{\numberline {3.2.4}Distributed data parallel}{51}{subsection.3.2.4}%
\contentsline {section}{\numberline {3.3}Google cloud compute instance }{52}{section.3.3}%
\contentsline {section}{\numberline {3.4}Repository del progetto}{52}{section.3.4}%
\contentsline {chapter}{\numberline {4}Sviluppo del progetto}{53}{chapter.4}%
\contentsline {section}{\numberline {4.1}Definizione della pipeline di addestramento}{53}{section.4.1}%
\contentsline {section}{\numberline {4.2}Preparazione dei dati per la pipeline}{55}{section.4.2}%
\contentsline {subsection}{\numberline {4.2.1}Severstal steel defect detection dataset reader}{55}{subsection.4.2.1}%
\contentsline {subsection}{\numberline {4.2.2}Conversione nel formato line json}{56}{subsection.4.2.2}%
\contentsline {subsection}{\numberline {4.2.3}Split in train e test set}{58}{subsection.4.2.3}%
\contentsline {subsection}{\numberline {4.2.4}Creazione del dataset degli oggetti}{59}{subsection.4.2.4}%
\contentsline {subsection}{\numberline {4.2.5}Creazione del dataset di immagini base}{60}{subsection.4.2.5}%
\contentsline {subsection}{\numberline {4.2.6}Creazione del dataset utilizzato come riferimento}{61}{subsection.4.2.6}%
\contentsline {section}{\numberline {4.3}Il dataloader}{61}{section.4.3}%
\contentsline {subsection}{\numberline {4.3.1}L'oggetto Augmentor}{61}{subsection.4.3.1}%
\contentsline {subsection}{\numberline {4.3.2}Gli oggetti JsonLineObjectDataset e JsonLineMaskObjectDataset}{63}{subsection.4.3.2}%
\contentsline {subsection}{\numberline {4.3.3}L'oggetto ObjectDataloader e ShapeObjectDataloader}{64}{subsection.4.3.3}%
\contentsline {subsection}{\numberline {4.3.4}L'oggetto ImageFolder}{64}{subsection.4.3.4}%
\contentsline {subsection}{\numberline {4.3.5}Il Noise Generator}{65}{subsection.4.3.5}%
\contentsline {subsection}{\numberline {4.3.6}L'oggetto CoiganSeverstalSteelDefectsDataset}{67}{subsection.4.3.6}%
\contentsline {section}{\numberline {4.4}La pipeline di addestramento}{68}{section.4.4}%
\contentsline {subsection}{\numberline {4.4.1}Il generatore}{68}{subsection.4.4.1}%
\contentsline {subsubsection}{La non saturating loss}{69}{section*.86}%
\contentsline {subsubsection}{La loss L1 smooth masked}{69}{section*.87}%
\contentsline {subsubsection}{La path-length regularization}{71}{section*.89}%
\contentsline {subsection}{\numberline {4.4.2}La perceptual loss smooth masked}{72}{subsection.4.4.2}%
\contentsline {subsection}{\numberline {4.4.3}Il discriminatore dei difetti}{74}{subsection.4.4.3}%
\contentsline {subsection}{\numberline {4.4.4}Il discriminatore di riferimento}{76}{subsection.4.4.4}%
\contentsline {section}{\numberline {4.5}La valutazione del modello}{77}{section.4.5}%
\contentsline {subsection}{\numberline {4.5.1}Fréchet Inception Distance FID}{78}{subsection.4.5.1}%
\contentsline {subsection}{\numberline {4.5.2}I risultati}{79}{subsection.4.5.2}%
\contentsline {chapter}{\numberline {5}Conclusioni e sviluppi futuri}{80}{chapter.5}%
\contentsline {section}{\numberline {5.1}Sviluppi futuri}{80}{section.5.1}%
\contentsline {chapter}{Elenco delle figure}{81}{chapter*.94}%
\contentsline {chapter}{Bibliografia}{86}{chapter*.95}%