Vein pattern visualisation using condition based transformer embedded generative adversarial networks (CTransGAN)
Loading...
Supplementary material
Other Title
Authors
Chhetri, Sumit
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2025
Supervisors
Sharifzadeh, Hamid
Keivanmarz, Ali
Keivanmarz, Ali
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
vein patterns
forensics
biometrics
child pornography
victim identification
criminal indentification
autoencoder
New Zealand
forensics
biometrics
child pornography
victim identification
criminal indentification
autoencoder
New Zealand
ANZSRC Field of Research Code (2020)
Citation
Chhetri, S. (2025). Vein pattern visualisation using condition based transformer embedded generative adversarial networks (CTransGAN) (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
https://hdl.handle.net/10652/6805
Abstract
With the increase in crimes such as child sexual abuse, identifying both the culprit and the victim has become more crucial than ever. Traditional biometrics such as fingerprint, facial pattern, and palmprint are often not applicable in these cases and are highly prone to forgery, reducing their reliability. Recently, researchers have turned to vein pattern recognition as a potential solution for subject identification. However, veins are nearly invisible to the human eyes and not easily observable in the RGB images taken from normal cameras. While earlier studies heavily relied on image quality, more recent research has predominantly utilised convolutional-based architectures. While these convolutional-based architectures are effective at capturing local spatial features, they struggle to capture spatial features for long-range dependencies. The veins in the forearm or hand generally have a long and elongated non-linear structure which spans over a larger area as compared to the finger or palm. Thus, the model’s capability to capture the long-range dependencies is critical in this situation.
A novel Transformer-encoder embedded conditional GAN (CTransGAN) model to address this limitation is proposed in this thesis. The proposed CTransGAN model is designed to generate NIR images from input RGB images. It incorporates three main components in the generator: an encoder and decoder based on a Convolutional Neural Network (CNN), along with a transformer module. The encoder extracts local features from the input RGB images while reducing their spatial dimensions. The Transformer module processes these features to capture global dependencies across the image, such as the distant or non-adjacent structure of the vein pattern. The decoder then reconstructs the output NIR images by upsampling and restoring the spatial details. Additionally, a CNN-based discriminator is employed, which concatenates both the input RGB image and the generated NIR image as input, helping it differentiate between target and output NIR images. This setup enables the model to effectively combine local and global feature representation for accurate vein pattern visualisation. A New Zealand-based RGB-NIR synchronised images dataset of 602 forearms with palms has been used. The dataset is processed and used to train and test the model in three distinct variations (100x100, forearms, and full hands). The evaluation uses metrics such as Structural Similarity Index Measure (SSIM), contrast accuracy, Peak Signal-to-Noise Ratio (PSNR), and vein length comparison upon the generated NIR image. Moreover, the model’s performance is further assessed through visual validation. The results of the proposed model surpass the performance of the existing state-of-the-art models used in this area of research.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved