/
11-00-ACA-Fingerprinting.tex
260 lines (239 loc) · 12.3 KB
/
11-00-ACA-Fingerprinting.tex
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
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
% move all configuration stuff into includes file so we can focus on the content
\input{include}
\subtitle{module 11.0: audio fingerprinting}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
% generate title page
\input{include/titlepage}
\section[overview]{lecture overview}
\begin{frame}{introduction}{overview}
\begin{block}{corresponding textbook section}
%\href{http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6331126}{Chapter 9: Audio Fingerprinting} (pp.~163--167)
chapter~11
\end{block}
\begin{itemize}
\item \textbf{lecture content}
\begin{itemize}
\item introduction to audio fingerprinting
\item in-depth example for fingerprint extraction and retrieval
\end{itemize}
\bigskip
\item<2-> \textbf{learning objectives}
\begin{itemize}
\item discuss goals and limitations of audio fingerprinting systems as compared to watermarking or cover song detection systems
\item describe the processing steps of the Philips fingerprinting system
\end{itemize}
\end{itemize}
\inserticon{directions}
\end{frame}
\section[intro]{introduction}
\begin{frame}{audio fingerprinting}{introduction}
\begin{itemize}
\item \textbf{objective}:
\begin{itemize}
\item represent a recording with a compact and unique digest\\ ($\rightarrow$ \textit{fingerprint}, \textit{perceptual hash})
\bigskip
\item<2-> allow quick matching between previously stored fingerprints and an extracted fingerprint
\end{itemize}
\bigskip
\item<3-> \textbf{applications}:
\begin{itemize}
\item \textit{broadcast monitoring}:\\ automate verification for royalties/infringement claims
\item \textit{value-added services}:\\ offer information and meta data
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{audio fingerprinting}{fingerprinting vs.\ watermarking}
\begin{itemize}
\item \textbf{fingerprinting}:
\begin{itemize}
\item identifies \textit{recording} (but not musical content)
\end{itemize}
\item \textbf{watermarking}:
\begin{itemize}
\item embeds perceptually ``unnoticeable'' data block in the audio
\item identifies \textit{instance} of recording
\end{itemize}
\end{itemize}
\pause
\begin{footnotesize}
\begin{table}
\centering
\begin{tabular}{lccccccccccc} %{\textwidth}{@{\extracolsep{\fill}}ccccccccccccc}
\\ \hline
\bf{\emph{Property}} & \bf{\emph{Fingerprinting}} & \bf{\emph{Watermarking}}\\
\hline
\bf{Allows Legacy Content Indexing} & + & --\\
\bf{Allows Embedded (Meta) Data} & -- & +\\
\bf{Leaves Signal Unchanged} & + & --\\
\bf{Identification of} & Recording & User or Interaction\\
\end{tabular}
\end{table}
\end{footnotesize}
\end{frame}
\section[requirements]{requirements}
\begin{frame}{audio fingerprinting}{fingerprint requirements}
\begin{itemize}%
\item \textbf{accuracy \& reliability}:\\ minimize false negatives/positives
\item<2-> \textbf{robustness \& security}: \\ robust against distortions and attacks
\item<3-> \textbf{granularity}:\\ quick identification in a real-time context
\item<4-> \textbf{versatility}:\\ independent of file format, etc.
\item<5-> \textbf{scalability}:\\ good database performance
\item<6-> \textbf{complexity}:\\ implementation possible on embedded devices
\end{itemize}
\end{frame}
\section[approaches]{approaches}
\begin{frame}{audio fingerprinting}{general fingerprinting system}
\begin{figure}
\centering
\input{pict/fingerprinting_system}
\end{figure}
\end{frame}
\begin{frame}{audio fingerprinting}{brainstorm}
\question{How does it work? MD5?}
\end{frame}
\section{philips}
\begin{frame}{audio fingerprinting}{system example: philips extraction 1/3}
\begin{footnotesize}
\begin{columns}[T]
\column{.3\textwidth}
\scalebox{.75}
{
\centering
\input{pict/fingerprinting_flowphilips}
}
\column{.7\textwidth}%\vspace{-5mm}
\begin{enumerate}
\item<1-> \textbf{pre-processing}:\\ downmixing \& downsampling (\unit[5]{kHz})
\item<2-> \textbf{STFT}: $\mathcal{K}=2048$, overlap $\frac{31}{32}$
\item<3-> \textbf{log frequency bands}:\\ $33$ bands from 300--2000\unit{Hz}
\item<4-> \textbf{freq derivative}: $33$ bands
\item<4-> \textbf{time derivative}: $32$ bands
\item<5-> \textbf{quantization}:
\begin{tiny}
\begin{equation*}\label{eq:fingerprint}
v_\mathrm{FP}(k,n) = \begin{cases}
1 & \text{if } \big(\Delta{E}(k,n) - \Delta{E}(k,n-1)\big) > 0\\
0 & \text{otherwise}
\end{cases}\nonumber
\end{equation*}
\end{tiny}
\item<6->[$\Rightarrow$] \textbf{\unit[32]{bit}} \textit{subfingerprint}
\end{enumerate}
\end{columns}
\end{footnotesize}
\end{frame}
\begin{frame}{audio fingerprinting}{system example: philips extraction 2/3}
\begin{itemize}
\item[] \textbf{fingerprint}
\begin{itemize}
\item $256$ subsequent subfingerprints
\item<2->[$\Rightarrow$]
\item<2-> \textit{length}: \unit[3]{s}
\item<2-> \textit{size}: $256\cdot \unit[4]{Byte} = \unit[1]{kByte}$
\end{itemize}
\smallskip
\item<3->[] \textbf{example}:
\begin{itemize}
\item \unit[5]{min} song
\begin{equation*}
\unit[1]{kByte} \cdot \frac{5\cdot 60 \unit{s}}{\unit[3]{s}} = \unit[100]{kByte}
\end{equation*}
\end{itemize}
\begin{itemize}
\item<4-> database with 1 million songs (avg.\ length \unit[5]{min})
\begin{equation*}
10^6\cdot 256\cdot\frac{5\cdot 60 \unit{s}}{\unit[3]{s}} = \unit[25.6\cdot 10^9]{subfingerprints}
\end{equation*}
\item<4->[$\Rightarrow$] \unit[100]{GByte} storage
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{audio fingerprinting}{system example: philips extraction 3/3}
\begin{columns}
\column{.4\linewidth}
\begin{itemize}
\item original: \includeaudio{pop_excerpt}
\item low quality encoding: \includeaudio{pop_excerpt_lq}
\end{itemize}
\column{.6\linewidth}
\vspace{-5mm}
\figwithmatlab{Fingerprint}
\end{columns}
\end{frame}
\begin{frame}{audio fingerprinting}{system example: philips identification 1/3}
\begin{itemize}
\item \textbf{database}
\begin{itemize}
\item contains all subfingerprints for all songs
\item<2-> previous example database: $25$ billion subfingerprints
\bigskip
\end{itemize}
\item<3-> \textbf{problem}
\begin{itemize}
\item how to identify fingerprint efficiently?
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{audio fingerprinting}{system example: philips identification 2/3}
\begin{itemize}
\item \textbf{simple system}:
\begin{enumerate}
\item create lookup table with all possible subfingerprints ($2^{32}$) pointing to occurrences
\bigskip
\item<2-> assume at least one of the extracted $256$ subfingerprints is error-free\\
\item<2->[$\Rightarrow$] only entries listed at $256$ positions of the table have to be checked
\bigskip
\item<3-> compute \textit{Hamming} distance between extracted fingerprint and candidates
\end{enumerate}
\end{itemize}
\end{frame}
\begin{frame}{audio fingerprinting}{system example: philips identification 3/3}
\begin{itemize}
\item \textbf{variant 1}:
\begin{itemize}
\item allow \textit{one} bit error
\item<2->[$\Rightarrow$] workload increase by factor $\approx 33$
\end{itemize}
\bigskip
\item<3-> \textbf{variant 2}:
\begin{itemize}
\item<3-> introduce concept of bit error probability into fingerprint extraction
\begin{itemize}
\item small energy difference $\rightarrow$ high error probability
\item large energy difference $\rightarrow$ low error probability
\end{itemize}
\bigskip
\item<4-> rank bits per subfingerprint by error probability and check only for bit errors at likely positions
\end{itemize}
\end{itemize}
\end{frame}
\section{shazam}
\begin{frame}{audio fingerprinting}{other systems: shazam}
\vspace{-3mm}
\begin{figure}
\centering
\includegraphics[scale=.23]{graph/fingerprint_shazaam}
\end{figure}
\vspace{-5mm}
plot from \footfullcite{wang_industrial_2003}
\end{frame}
\section{summary}
\begin{frame}{summary}{lecture content}
\begin{itemize}
\item \textbf{audio fingerprinting}
\begin{itemize}
\item represent recording with compact, robust, and unique fingerprint
\item focus on (perceptual) audio representation rather than ``musical'' content
\item allow efficient matching of this fingerprint with database
\end{itemize}
\bigskip
\item \textbf{often confused with other tasks}
\begin{enumerate}
\item audio watermarking
\item cover song detection
\end{enumerate}
\end{itemize}
\inserticon{summary}
\end{frame}
\end{document}