Skip to content
Browse files

lot of fixes

  • Loading branch information...
1 parent f4eacc3 commit 15d3125f3597fef8b9bef57c51de9da4118af85b neuromancer committed Jul 29, 2012
Showing with 5 additions and 8 deletions.
  1. +5 −8 baravalle.tex
View
13 baravalle.tex
@@ -179,11 +179,9 @@ \subsubsection*{Procedure}
\section{Materials and Methodology}
\subsection{Image Acquisition}
-Fifty images of four different bread types (lactal, baguette, salvado and sandwich), counting two hundred ($200$) images, were obtained using an electric slicer (in the case of baguette and salvado, since the two other types were already sliced in the moment of purchase). The images were digitalised using an HP PSC 1210 scanner and they were saved in TIFF format. Images showed a resolution of $380 \times 380$ pixels (the maximum possible surface for the four types of bread) and $350$ dpi ($1$ pixel - $0.00527 mm^{2}$). Then they were converted to grey scale ($8$ bits). Also $20$ images of each bread type were taken with a digital camera, using the same spatial resolution, counting $80$ images. The illumination conditions of these images were different from that of the scanner in order to test for the robustness of the method. We also employed images from the dataset CalTech101~\cite{FeiFei04} in order to test the method's performance with non-bread images.
+Fifty images of four different bread types (lactal, baguette, salvado and sandwich), counting two hundred ($200$) images, were obtained using an electric slicer (in the case of baguette and salvado, since the two other types were already sliced in the moment of purchase). The images were digitalised using an HP PSC 1210 scanner and they were saved in TIFF format. Images showed a resolution of $380 \times 380$ pixels (the maximum possible surface for the four types of bread) and $350$ dpi ($1$ pixel - $0.00527 mm^{2}$). Then they were converted to grey scale ($8$ bits). Also $20$ images of each bread type were taken with a digital camera, using the same spatial resolution, counting $80$ images. The illumination conditions of these images were different from that of the scanner in order to test for the robustness of the method. In Fig. \ref{fig:camera} four examples of bread images from the camera are shown. We also employed 20 \todo{Check} randomly selected images from the dataset CalTech101~\cite{FeiFei04} in order to test the method's performance with non-bread images. In Fig.~\ref{fig:nonbread} four examples of non-bread images from this dataset are shown.
-In those cases where the utilised procedure uses a binarisation of the original image, it was obtained using the White's algotithm \cite{White83}. This algorithm applies a local thresholding schema, which showed better results in comparison to those obtained using a global thresholding schema. Particularly, the algorithm presented in \cite{Huang95} and used in \cite{Gonzales2008}, showed poor results when the illumination conditions varies. Since the center of air bubbles with bigger areas appeared as black pixels, instead of white (and since those areas are characterized as dark regions in the original image), a global grey threshold using Otsu's algorithm \cite{Otsu79} is obtained and multiplied with a scalar which is a parameter, setting as white the resulting pixels in the final binarisation. It was found that defining the scalar as $0.8$ showed acceptable results. So the combination of local and global thresholding makes it an hybrid algorithm.
-
-In Fig. \ref{fig:bread} an image of each bread type used in this work and its resulting binarisaton using the proposed algorithm is shown. In Fig. \ref{fig:nonbread} four examples of non-bread images from the dataset CalTech101 are shown. The classificator is expected to classify them as non-bread type. In Fig. \ref{fig:camera} four examples of bread images from the camera are shown.
+In those cases where the utilised procedure uses a binarisation of the original image, it was obtained using the White's algotithm \cite{White83}. This algorithm applies a local thresholding schema, which showed better results in comparison to those obtained using a global thresholding schema. Particularly, the algorithm presented in \cite{Huang95} and used in \cite{Gonzales2008}, showed poor results when the illumination conditions varies. Since the center of air bubbles with bigger areas appeared as black pixels, instead of white (and since those areas are characterized as dark regions in the original image), a global grey threshold using Otsu's algorithm \cite{Otsu79} is obtained and multiplied with a scalar which is a parameter, setting as white the resulting pixels in the final binarisation. It was found that defining the scalar as $0.8$ showed acceptable results. So the combination of local and global thresholding makes it an hybrid algorithm. In Fig.~\ref{fig:bread} an image of each bread type used in this work and its resulting binarisaton using the proposed algorithm is shown.
\begin{figure*}[htb]
@@ -223,12 +221,11 @@ \subsection{Image Acquisition}
\subsection{Feature Vectors}
-Following the ideas presented in \cite{Gonzales2008}, the mentioned fractal and multifractal features were obtained for each image (using $20$ H\"older exponents, a $42$ features vector was obtained). The code of the algorithms Box dimension, Morphological fractal dimension and the multifractal spectrum was written and run in Matlab. In order to make a comparison, a vector with RGB color features was used (R mean, G mean, B mean), computing a $3$ dimension features vector.
+Following the ideas presented in \cite{Gonzales2008}, the mentioned fractal and multifractal features were obtained for each image (using $20$ H\"older exponents). For each image, a $42$ features vector was computed. The code of the algorithms Box dimension, Morphological fractal dimension and the multifractal spectrum was written and run in Matlab. In order to make a comparison, a vector with RGB color features was computed (R mean, G mean, B mean) in a $3$ dimension features vector.
-\subsection{Self-Organizing Maps (SOM)}
-SOM\todo{ Add cite Kohonen} are non-supervised generated maps. A SOM maps high dimensional data into a (tipically) two-dimensional representation, using neighborhood information, so preserving topological information of the original data. They are useful to visualize the input data into a lower dimensional view, in order to better understand it, and analyse its spatial representation.
+A self-organizing maps (SOM)~\cite{Kohonen2001} of the vectorized images are useful to visualize these different representation of bread images into a lower dimensional view, in order to understand them better. A SOM maps high dimensional data into a (tipically) two-dimensional representation, using neighborhood information. Topological information of the original data is preserved.
-An unsupervised Self Organizing Map (SOM) of the fractal and non fractal representation of scanned images are shown in Fig. \ref{fig:somfractal} and \ref{fig:somrgb} respectively. On the one hand, the fractal features SOM seems to show easily separable classes, on the other hand, the RGB features appears to be more overlapped. Also, in the latter, the non bread class seems to be spread over the rest of the classes, making it more difficult to distinguish between bread and non bread types. It seems that a classificator could get better classification results using the fractal features. Next sections will show that this hypothesis is true.
+An unsupervised SOM of the fractal and non fractal representation of scanned images are shown in Fig.~\ref{fig:somfractal} and Fig.~\ref{fig:somrgb} respectively. The fractal features SOM seems to show easily separable classes and the RGB features SOM appears to be more overlapped. Also, in the latter, the non bread class seems to be spread over the rest of the classes, making it more difficult to distinguish between bread and non bread types. It seems that a classificator could potentialy obtain better classification results using the fractal features. %%Next sections will show that this hypothesis is true.
\begin{figure*}[]
\centering

0 comments on commit 15d3125

Please sign in to comment.
Something went wrong with that request. Please try again.