Vein patterns are gaining attention as a biometric modality, particularly in forensic investigations involving child sexual abuse cases where other biometric traits are often unavailable. Despite their strength as an identification method, extracting vein patterns from standard RGB images remains a challenging task. Although various techniques have been proposed to visualise veins from colour images, most predominantly rely on convolutional architectures. While these models effectively capture fine-grained local pixel details, they often struggle to preserve long-range dependencies and broader spatial context. However, when dealing with complex vein structures, intricate branching patterns, and extended regions, capturing global dependencies becomes essential for accurate visualisation.In this paper, we propose a hybrid Transformer Encoder Embedded Conditional Generative Adversarial Network (CTrans-GAN), which integrates the strengths of both convolutional networks and transformers through a self-attention mechanism. The model is trained on an RGB-NIR paired image dataset from 301 subjects and evaluated on arm images across three variations: small 100×100 crops, forearm regions, and full arm images. The performance of the proposed model is assessed using a range of objective metrics, including contrast accuracy, PSNR, SSIM, and vein length accuracy. The evaluation results demonstrate that the proposed model not only outperforms the previous GAN model on the 100×100 cropped images but also excels in visualising veins across larger skin areas, including the forearm and full arm, achieving over 81% vein length accuracy across all three dataset variations.

Traffic signal control is the key to managing urban traffic flows and volumes. However, traditional traffic signal control methods typically struggle to deal with real-time traffic fluctuations, leading to frequent congestion. In contrast, reinforcement learning (RL), a machine learning approach that learns and optimises behaviour from its environment, offers an adaptive way to adjust signal control strategies. This paper proposes a new RL method based on Enhanced Deep Q-networks (EDQN) to optimise traffic signal phasing at a single intersection, giving additional neural networks to reduce further overestimation on Q-values compared with other methods. To demonstrate the optimisation of the enhanced model, we define the average queue length of all vehicles that wait in the lane before entering the intersection as the evaluation metric. The results of simulation with light, medium and heavy traffic volumes show that this model achieves over 70% optimisation in reducing queue length.

Road surface damage detection is important for maintaining transportation infrastructure and ensuring public safety. Automated approaches help reduce labour-intensive manual inspections and expedite critical maintenance decisions. In this paper, we propose an efficient, two-stage object detection framework that balances speed and accuracy, using Faster R-CNN as its core. The novelty of our approach lies in custom enhancements to the Faster R-CNN architecture, improving detection precision while preserving computational feasibility.To validate this framework, we benchmark multiple object detection models, including the default Faster R-CNN, Sparse R-CNN, YOLOv5L, YOLOv8L, YOLOv11, a customised sparsified Faster R-CNN, and the optimised Faster R-CNN. The models are trained and tested on the RDD2022 dataset with over 47000 images capturing 9 categories of road surface damages. Results show that our optimised Faster R-CNN achieves over 94% alignment between predicted bounding boxes and ground truth annotations, outperforming other models in precision. Although it operates at a slower rate of 11 images per second compared to YOLOv5L’s 52 images per second, its higher accuracy makes it especially valuable for high-stakes applications such as automated pavement distress assessment.