Arani Aravinthan and Madushi Pathmaperuma, University of Central Lancashire, UCL
Beauty is a power which allows people to express themselves, gain self-confidence and open to others. Usage of beauty products can help this by creating a new look to uplift the character. Choosing the right makeup product is not an easy task in this diverse range of products these days. Intelligent systems for beauty and makeup selection have gained significant research interest in recent years. Most existing models focus on detecting prominent facial features such as skin tone, lip colour, and overall facial structure. However, minor yet impactful areas, such as the delicate regions around the eyes, are often overlooked. These areas play a crucial role in defining facial aesthetics, influencing expressions, and enhancing overall appearance. To address this gap, this system is designed to provide targeted recommendations for eye-focused beauty enhancements, ensuring a more comprehensive and personalized approach to makeup selection. The proposed system will recommend makeup products considering personal traits of the user such as the length and volume of the eyelashes. A new approach has been devised in calculating the length of the eyelashes aiding the use of advanced computer vision techniques like edge detection, and a regression-based Convolutional Neural Network (CNN) model is trained for prediction. A Support Vector Machine (SVM) is used for the classification task in recommending products for eyelash care.
Edge Detection, Contour Detection, Support Vector Machine.
Massimiliano Barone, STMicroelectronics, Agrate Brianza, Milano, Italy
The goal of this paper is to present a new, fast, and sub-optimal technique for splitting overly large regions of interest (ROIs) or the original full image. ROIs are typically box areas containing significant image information. The aim is to generate ROIs based on a specific criterion for subsequent focused analysis. The proposed technique converts a generic ROI size or the original full image into a few smaller, fixed-size ROIs, fully covering all significant details, with a maximum interception over union and a high number of details for each new ROI. Where details typically come from a previous feature extraction or image segmentation. This method is particularly helpful for convolutional neural networks (CNN) with fixed-size input images and datasets requiring image portions with as much pertinent information as possible. The effectiveness is compared against dividing the original ROI into a regular grid of ROIs, using the number of ROIs and the average detail density per ROI.
Regions of interest, ROI, clustering, segments, computer vision, regional proposal.
Atoshe Islam Sumaya, M. Nessa, and T. Rahman, School of Engineering, BRAC University, Bangladesh
FIR Filter is widely used for image processing, wireless communication, radar systems, control systems, and biomedical signal processing for its intrinsic stability and linear phased attribute. This design implements a prime number coefficient value 41 to decrease periodic faults together with symmetric arti-facts that appear in the frequency response. The selected number achieves an optimal relationship between processing speed and signal quality. Which makes it appropriate for real-time and high-performance sys-tems. The design uses Skywater 130nm CMOS technology and an ensemble of open-source tools starting from RTL and ending at GDSII production. These tools include Openlane, OpenROAD, Magic, Klayout, Netgen, and Yosys. It also includes industry-standard DRC and LVS checks for verification. The imple- mented layout contains 642,151 square micrometers of space with 72.61 MHz operation speed and 0.266 microwatt power utilization. The proposed 41-tap fir filter design introduces significant improvement in performance and power consumption compared to existing work.
Sky130 PDK, Heatmap, OpenLane, Magic, Klayout.
Alexander Dylan Bodner, Jack Natan Spolski ,Antonio Santiago Tepsich, Santiago Pourteau ,Universidad de San Andres ,Argentina
In this paper, we present Convolutional Kolmogorov-Arnold Networks, a novel architecture that integrates the learnable spline-based activation functions of Kolmogorov-Arnold Networks (KANs) into convolutional layers. By replacing traditional fixed-weight kernels with learnable non-linear functions, Convolutional KANs offer a significant improvement in parameter efficiency and expressive power over standard Convolutional Neural Networks (CNNs). We empirically evaluate Convolutional KANs on the Fashion-MNIST dataset, demonstrating competitive accuracy with up to 50% fewer parameters compared to baseline classic convolutions. This suggests that the KAN Convolution can effectively capture complex spatial relationships with fewer resources, offering a promising alternative for parameter-efficient deep learning models.
Machine Learning, Kolomogorov-Arnold Networks, Convolutional Kolmogorov-Arnold Networks.
Sun Xiaoping and Rodolfo C. Raga Jr, National University, Philippines
This paper proposes an air quality prediction model ST-HyDM based on spatiotemporal data mining and hybrid deep learning. By constructing a dynamic graph neural network to capture spatiotemporal dependencies, fusing multimodal features, and using a dual-branch architecture of Transformer and TCN for feature extraction and prediction. At the same time, a missing value processing mechanism based on generative adversarial networks (GAN) is introduced to improve data integrity and model performance. Experimental results show that compared with traditional models such as linear regression and random forest, ST-HyDM has higher accuracy and robustness in air quality prediction, providing a more effective solution for air quality prediction.
Air quality prediction, spatiotemporal data mining, hybrid deep learning, dynamic graph neural network, missing value processing.
Nuwan Kaluarachchi1, Arathi Arakala1, Sevvandi Kandanaarachchi2 and Kristen Moore2, 1School of Mathematical and Geospatial Sciences, RMIT University, Melbourne, Australia, 2CSIRO’s Data61, Melbourne, Australia
Keystroke dynamics is a behavioural biometric modality that utilises individual typing patterns for user authentication. While it has been popular on single device authentication, its application in cross-device scenarios remains under-explored. This paper proposes a solution for keystroke dynamics for cross-device user authentication using a transfer learning framework. Specifically, we authenticate users on tablets by training the authentication unit mostly on smartphone keystroke dynamics. We call our framework TEDxBC, as it includes a Transfer Encoder, a Data-fusion module, and a Binary Classifier for cross-device scenarios. Leveraging 24 keystroke dynamic features incorporating spatial and traditional features, TEDxBC employs an inductive transfer encoder to map users from the smartphone to the tablet. We evaluate TEDxBC on participants from the publicly available BBMAS dataset. Our method achieves an average Equal Error Rate (EER) of 14% , surpassing state-of-the-art methods on the same database. Furthermore, we apply a biometric menagerie analysis to gain insights into the performance of different user groups. Our analysis reveals that users in the “doves” group that authenticate with high accuracy on a single device retain their high performance in a cross-device scenario. They achieve an EER of 5% , surpassing the overall TEDxBC performance.
Authentication, Transfer learning, Cross-device biometric authentication, Keystroke dynam- ics, biometric menagerie.
Doliana Celaj1 and Greta Jani2, 1Bayswater College, England, 2Department of Albanian Languages, Faculty of Education, “Aleksander Moisiu” University, Durres, Albania
This study explores the impact of Natural Language Processing (NLP) tools on speaking and vocabulary development among ESL learners. Over a 6-week intervention, two groups of students are assessed—one using traditional learning methods, and the other supported by NLP-enhanced tools. The results revealed significant improvements in vocabulary retention, speaking performance, and learner confidence in the NLP-supported group, underscoring the value of technology-driven approaches in ESL instruction..
Natural Language Processing (NLP), ESL Education, Speaking Skills, Vocabulary Acquisition, Speech Recognition, Pronunciation Training, Student Motivation.
Copyright © NLPA 2025