design.tex

图像中非刚性曲线的蛇形检测算法

\section{Design of the GSNAKE API}
The GSNAKE API is a set of fundamental codes, based on the Generalised 
Active Contour Model \cite{kn:thesis}, suitable for use in the area of feature extraction, image detection and classification, and motion analysis. 

The API consists of a set of objects built in C++. These are described in the following section.  


\section{Class Design and Implementation}

\subsubsection*{class IMAGE}
{\tt IMAGE} provides routines for manipulating and displaying image data. These includes histogram conditioning, image correlation and image smoothing. 


\subsection*{class EDGE}
{\tt EDGE} computes edge map of an image. It consists of two {\tt IMAGE} objects that marking the magnitude and direction of each edge point.


\subsection*{class PYRAMID}
{\tt PYRAMID} is built up of a series of {\tt EDGE} and {\tt IMAGE} objects in a pyramid form. The higher level {\tt IMAGE} is a reduced version of the lower image in that both resolution and sample density are decreased. By computing and conditioning edge map from each level of {\tt IMAGE}, we obtain {\tt EDGE} objects.


\subsection*{class MATRIX}
{\tt MATRIX} provides advanced routines for matrix addition, substraction, multiplication, inversion and transpose. It is inherited with extra members and methods from {\tt IMAGE class}.


\subsection*{class SNAXEL}
A {\tt SNAXEL} is a point on a contour. It consists of methods for snaxel angle and energy calculation.


\subsection*{class CONTOUR}
{\tt CONTOUR} is a linked list of {\tt SNAXEL} originated at image origin or contour center. It consists of shape matrix and internal energy term to represent shape irregularities.


\subsection*{class GHOUGH}
{\tt GHOUGH} localizes {\tt CONTOUR} of a particular shape from an {\tt IMAGE} or {\tt EDGE} by generalized Hough transform.

\subsection*{class GSNAKE}
{\tt GSNAKE} consists of {\tt PYRAMID} for external energy calculation and {\tt CONTOUR} for internal energy calculation. It localizes {\tt CONTOUR} by generalized Hough transform, minimizes energy by dynamic programming with stratified line search algorithm, and regularizes parameters tradeoff by minmax criterion or local/global parameters selection strategy.

\subsubsection*{class MODEL}
With sufficient image samples, {\tt MODEL} trains a robust {\tt CONTOUR} model with prior knowledge of shape matrix and local regularization parameters. It is inherited with extra members and methods from {\tt GSNAKE}.

\subsubsection*{class CLASSIFY}
{\tt CLASSIFY} classifies various {\tt CONTOUR} models and compute their corresponding score from an image. Classifying methods include marginalization of the distribution, MAP probability, match of deformable template and rigid template.