\relax 
\bibstyle{plainnat}
\newlabel{sec:intro}{{1}{155}}
\@writefile{toc}{\contentsline {section}{\numberline {1}Introduction}{155}}
\citation{Moya2}
\citation{Vapnik}
\citation{Tax5}
\citation{Scholkopf3}
\citation{Scholkopf5}
\citation{Burges}
\citation{Vapnik}
\citation{Burges}
\newlabel{sec:SVDD}{{2}{156}}
\@writefile{toc}{\contentsline {section}{\numberline {2}Support vector data description}{156}}
\citation{Scholkopf3}
\citation{Bishop}
\citation{Vapnik}
\citation{Tax5}
\newlabel{eq:error}{{1}{157}}
\newlabel{eq:constr}{{2}{157}}
\newlabel{eq:L}{{3}{157}}
\newlabel{eq:finalconstr}{{4}{157}}
\newlabel{eq:orgL}{{5}{157}}
\citation{Vapnik}
\citation{Tax9}
\citation{Tax5}
\citation{Tax5}
\citation{Scholkopf1}
\newlabel{eq:err_est}{{6}{158}}
\newlabel{eq:newtestz}{{7}{158}}
\newlabel{eq:gaussK}{{8}{158}}
\newlabel{eq:gaussL}{{9}{158}}
\newlabel{eq:finaltestz}{{10}{158}}
\newlabel{eq:def_nu}{{11}{158}}
\citation{Tax5}
\@writefile{toc}{\contentsline {section}{\numberline {3}Optimization of $s$ and $\nu $}{159}}
\newlabel{sec:opt_s_c}{{3}{159}}
\@writefile{lof}{\contentsline {figure}{\numberline {1}{\ignorespaces Decision boundaries of a simple 2-dimensional dataset, with different choices of $s$ and $\nu $.}}{159}}
\newlabel{fig:all_s_C}{{1}{159}}
\newlabel{eq:tot_error}{{12}{160}}
\citation{Luban}
\citation{Ullman}
\@writefile{toc}{\contentsline {section}{\numberline {4}Generating samples uniformly from a hypersphere}{161}}
\newlabel{sec:gen_sphere}{{4}{161}}
\newlabel{sec:exp}{{5}{162}}
\@writefile{toc}{\contentsline {section}{\numberline {5}Experiments }{162}}
\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}Artificial example}{162}}
\@writefile{lof}{\contentsline {figure}{\numberline {2}{\ignorespaces Left: Artificial outlier data generated around the banana-shaped target set from Figure 1\hbox {}. Right: SVDD decision boundary for optimized $s$ and $\nu $.}}{162}}
\newlabel{fig:opt_s_c}{{2}{162}}
\citation{Cho}
\@writefile{toc}{\contentsline {subsection}{\numberline {5.2}Handwritten digits}{163}}
\@writefile{lof}{\contentsline {figure}{\numberline {3}{\ignorespaces Testing target objects rejected by the data description.}}{163}}
\newlabel{fig:con_tst_3}{{3}{163}}
\@writefile{lof}{\contentsline {figure}{\numberline {4}{\ignorespaces Artificial objects (images of handwritten threes) generated from a box-shaped distribution around the target set (left subplot) and from a spherical-shaped distribution (right subplot).}}{164}}
\newlabel{fig:con_reconstr}{{4}{164}}
\@writefile{toc}{\contentsline {subsection}{\numberline {5.3}Texture segmentation}{164}}
\@writefile{lof}{\contentsline {figure}{\numberline {5}{\ignorespaces Texture segmentation. The texture in the cross shaped region in the middle is used as target class (left figure). The next three images are examples of textures used.}}{164}}
\newlabel{fig:texture_seg}{{5}{164}}
\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces Performances on image \emph  {1} using a trainingset of 400 patches (size $16\times 16$). 160 features are used. Results are averaged over 5 runs and multiplied by $100$. Numbers in brackets give the standard deviation.}}{165}}
\newlabel{tab:1_160D}{{1}{165}}
\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces Performances on image \emph  {1} using a trainingset of 400 patches. 35 features are used. Results are averaged over 5 runs and multiplied by $100$. Numbers in brackets give the standard deviation.}}{167}}
\newlabel{tab:1_35D}{{2}{167}}
\@writefile{lot}{\contentsline {table}{\numberline {3}{\ignorespaces Performances on image \emph  {1} using a trainingset of 400 patches. 12 features are used. Results are averaged over 5 runs and multiplied by $100$. Numbers in brackets give the standard deviation.}}{167}}
\newlabel{tab:1_12D}{{3}{167}}
\@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces Performances on image \emph  {1} using a trainingset of 400 patches. Outliers are generated in a hyperbox, instead of in a hypersphere. 12 features are used. Results are averaged over 5 runs and multiplied by $100$. Numbers in brackets give the standard deviation.}}{168}}
\newlabel{tab:1_12D_box}{{4}{168}}
\@writefile{lot}{\contentsline {table}{\numberline {5}{\ignorespaces Performances on image \emph  {2} using a trainingset of 400 patches (size $16\times 16$). 10 features are used. Results are averaged over 5 runs and multiplied by $100$. Numbers in brackets give the standard deviation.}}{169}}
\newlabel{tab:2_10D}{{5}{169}}
\@writefile{lof}{\contentsline {figure}{\numberline {6}{\ignorespaces Performance on the target and outlier set, \ensuremath  {f_{T-}}\xspace  and \ensuremath  {f_{O+}}\xspace  , for varying $\lambda $. Left the results for image \emph  {1} is shown, right for image \emph  {3}. Results are averaged over 5 runs.}}{169}}
\newlabel{fig:lamb_con_n_1_13}{{6}{169}}
\citation{Bishop}
\citation{Dempster}
\citation{akaike}
\@writefile{lof}{\contentsline {figure}{\numberline {7}{\ignorespaces Output of SVDD with optimized $s$ and $\nu $ for (left) image \emph  {1} and (right) image \emph  {3}, $\lambda =\frac  {1}{2}$. A light grayvalue indicates that the pixel is near the center ${\ensuremath  {\bf  a }}$ of the SVDD hypersphere (and thus classified as target object by the SVDD).}}{170}}
\newlabel{fig:im_1_13}{{7}{170}}
\citation{Vapnik}
\@writefile{lot}{\contentsline {table}{\numberline {6}{\ignorespaces Performances of the Mixture of Gaussians (with different number of clusters, $k$) compared with the SVDD on the three images. A trainingset of 400 patches is used (size $16\times 16$). Results are averaged over 5 runs and multiplied by $100$. Numbers in brackets give the standard deviation.}}{171}}
\newlabel{tab:res_mog}{{6}{171}}
\newlabel{sec:conclusions}{{6}{171}}
\@writefile{toc}{\contentsline {section}{\numberline {6}Conclusion }{171}}
\bibdata{/home/davidt/tex/bib/refs}