[WIP] add docs for data splitting and evaluation setups

struggling-student · Jan 12, 2025 · cc641a8 · cc641a8
1 parent f2a071c
commit cc641a8
Show file tree

Hide file tree

Showing 6 changed files with 44 additions and 10 deletions.
diff --git a/docs/src/abstract.tex b/docs/src/abstract.tex
@@ -1,3  1,3 @@
 \begin{abstract}
-    VeinNet is a Convolutional Neural Network (CNN)–based system for palm vein identification. The approach begins by extracting the palm region of interest (ROI) from a dataset of multispectral hand images, focusing on the vein patterns within this ROI as input to the CNN. Trained on these hand images, VeinNet demonstrates the ability to effectively recognize unique vein patterns in the human palm. The system achieves a xx\% accuracy on the test dataset, showcasing the effectiveness of a CNN-based solution.
     VeinNet is a Convolutional Neural Network (CNN)–based system for palm vein identification. The approach begins by extracting the palm region of interest (ROI) from a dataset of multispectral hand images, focusing on the vein patterns within this ROI as input to the CNN. Trained on these hand images, VeinNet demonstrates the ability to effectively recognize unique vein patterns in the human palm. The system is evaluated using three distinct types of evaluations: identification with a closed set, identification with an open set, and verification. The dataset is split differently for each evaluation setup, ensuring that the system is tested under various conditions. The results demonstrate the effectiveness of VeinNet in identifying individuals based on their palm vein patterns, with promising performance metrics across all evaluation setups.
 \end{abstract}
diff --git a/docs/src/dataset.tex b/docs/src/dataset.tex
@@ -1,22  1,22 @@
 \section{Dataset}
 This section provides an overview of the CASIA Multi-Spectral Palmprint Image Database, including data acquisition details and a description of the dataset.
 
 \subsection{Data Aquisition}
 
-The self-designed imaging device for acquiring hand images \cite{hao2008multispectral,hao2007comparative} is shown in Figure \ref{fig:device_architecture}. The device operates in a contact-free environment, during the imaging process:
 The self-designed imaging device for acquiring hand images \cite{hao2008multispectral,hao2007comparative} is shown in Figure \ref{fig:device_architecture}. The device operates in a contact-free environment, The imaging process involves the following steps:
 
 \begin{enumerate}
     \item Illumination Setup:
         \begin{itemize}
-            \item The device uses six groups of LEDs with wavelengths ranging from violet to near-infrared. These LEDs are turned on sequentially, allowing a time-division strategy for acquiring multispectral images.
-            \item This setup ensures the acquisition of images under varying illumination conditions, covering different layers of the skin due to light absorption and scattering properties.
             \item The device uses six groups of LEDs (violet to near-infrared) activated sequentially, employing a time-division strategy to acquire multispectral images under varying illumination, capturing different skin layers through light absorption and scattering. 
         \end{itemize}
     \item Reflective Imaging:
         \begin{itemize}
-            \item Images are captured in a reflective manner under a sheltered environment, ensuring consistent illumination and reducing external noise.
-            \item Each group of LEDs is arranged circularly and diffused using a ground glass to provide even illumination across the hand.
             \item Images are captured reflectively in a sheltered environment with consistent illumination, while circularly arranged LED groups, diffused with ground glass, ensure even lighting across the hand. 
         \end{itemize}
     \item Contact-Free Operation:
         \begin{itemize}
-            \item Subjects are instructed to naturally stretch their hands, palms facing the camera, without any physical contact with a tangible surface or plate. This setup enhances hygiene and minimizes user resistance.
             \item Subjects are instructed to naturally stretch their hands, palms facing the camera, without any physical contact with a tangible surface or plate. 
         \end{itemize}
     \item Sequential Image Capture:
         \begin{itemize}
@@ -28,7  28,7 @@ \subsection{Data Aquisition}
 
 \begin{figure}[H]
     \centering
-    \includegraphics[width=0.5\textwidth]{./images/device-architecture.png}
     \includegraphics[width=0.3\textwidth]{./images/device-architecture.png}
     \caption{Imaging Device Architecture for Hand Image Acquisition.}
     \label{fig:device_architecture}
 \end{figure}
@@ -39,7  39,7 @@ \subsection{Dataset Description}
 
 \begin{figure}[H]
     \centering
-    \includegraphics[width=0.5\textwidth]{./images/spectrums.png}
     \includegraphics[width=0.3\textwidth]{./images/spectrums.png}
     \caption{Palmprint images from the CASIA Multi-Spectral Palmprint Image Database with the six spectral bands. Starting from the top-left corner and moving clockwise: 460 nm, 630 nm, 700 nm, 850 nm, 940 nm, and white light.}
     \label{fig:dataset_example}
 \end{figure}
diff --git a/docs/src/discussion.tex b/docs/src/discussion.tex
diff --git a/docs/src/experimental.tex b/docs/src/experimental.tex
@@ -0,0  1,24 @@
 \section{Experimental}
 This section describes the experimental setup introducing the 3 evaluation setups used to assess the performance of the proposed biometric system and the data splitting strategy used for each evaluation setup.
 
 \subsection{Evaluation Setups}
 The evaluation was performed using three distinct types of evaluations:
 
 \begin{itemize}
     \item \textbf{Identification with a Closed Set}: involves including all enrolled patients in the dataset. Each image is classified into one of the known classes (patients) based on its extracted features. The system is trained and tested with the same predefined set of enrolled patients, ensuring no unknown users are present in the dataset.
 
     \item \textbf{Identification with an Open Set}: excludes a percentage of the enrolled patients from the dataset during training. These excluded patients represent unknown users during the evaluation phase. The model is tested on both known and unknown classes, where the unknown classes are expected to be classified as unknown to simulate open-set identification.
 
     \item \textbf{Verification}: tests the system's ability to verify the identity of users. Genuine samples consist of images correctly matched to their claimed identities, while imposter samples are created by associating images with incorrect user identities to simulate attempts to mislead the system.
 \end{itemize}
 
 \subsection{Data Splitting}
 The dataset used in this study was split differently for the three evaluation setups:
 
 \textbf{Identification with a Closed Set} includes all patients in the dataset, and their images are divided into training and test sets. The first four images per patient are used for training and the remaining images are used for testing. This setup ensures that all enrolled patients contribute images to the training and testing phases, facilitating evaluation in a controlled, closed-set scenario.
 
 \textbf{Identification with an Open Set} For open-set identification, 70\% of patients are randomly selected as known, with their images split into training (first four images) and testing (remaining images). The remaining 30\% serve as unknown patients, with their images used solely for testing to assess the system's ability to handle unenrolled users, simulating open-set scenarios. 
 
 \textbf{Verification} includes genuine and imposter samples for evaluation. For each patient, the first four images are used for training, and the remaining images are used for testing. Genuine samples consist of matching the correct image to the claimed identity, while imposter samples are created by pairing images from different patients, simulating attempts to impersonate other users. 
 
 To ensure consistency and reproducibility across all splits, a fixed random seed was used during shuffling. This guarantees that the data partitioning remains consistent across different runs and experimentsss
diff --git a/docs/src/methodology.tex b/docs/src/methodology.tex
@@ -1,6  1,7 @@
 \section{Methodology}
 
-\subsection{Data preprocessing}
 
 \subsection{Data Preprocessing}
 
 
 \subsection{Model Architecture}
diff --git a/docs/src/results.tex b/docs/src/results.tex
@@ -0,0  1,9 @@
 \section{Evaluation Metrics}
 
 \section{Results}
 
 \subsection{Identification with a Closed Set}
 
 \subsection{Identification with an Open Set}
 
 \subsection{Verification}