This paper proposes an innovative framework termed as FedCSG which has been designed to enhance graph learning on resource-limited devices. By leveraging federated learning and split learning techniques, FedCSG enables devices to collaboratively train models without sharing raw data, ensuring privacy and security. The novelty of the continual learning aspect allows the framework to adapt to new data over time, maintaining high performance even as the underlying data distribution changes. This makes FedCSG particularly suitable for applications in dynamic environments where devices have limited computational resources and need to process graph-structured data efficiently. The performance effectiveness of FedCSG is validated on graph datasets and experiments show that the proposed framework is superior to other state-of-the-art (SOTA) approaches.

The rapid proliferation of Artificial Intelligence of Things (AIoT) devices in smart cities, such as roadside sensors and traffic cameras, enables real-time urban traffic monitoring through distributed sensor networks. These systems generate spatio-temporal data critical for Intelligent Transportation Systems (ITS), particularly traffic forecasting, and prediction of congestion, accidents, and travel times. However, existing forecasting methods struggle with cross-city collaboration due to heterogeneous sensor topologies and privacy constraints, limiting their effectiveness. Federated Learning partially mitigates privacy issues but fails to effectively address the topological heterogeneity of sensor networks, impeding robust cross-city collaboration and generalization. To address these limitations, we propose FedGraphX, a Split Federated Graph Learning framework that integrates localized processing with global collaboration. Each city operates as an independent AIoT node, using GRU-based encoders and GraphSAGE models to extract spatio-temporal features from local data. A central Graph Transformer aggregates features across cities, linking sensors via temporal and functional similarities, while a cross-layer attention mechanism aligns local spatial patterns with global dynamics. Experiments on four real-world datasets demonstrate superiority of FedGraphX over centralized and federated baselines, particularly in scenarios with sparse data or topological mismatches.

Wireless rechargeable sensor networks (WRSNs) have emerged with strong potential to address the bottlenecks of energy/lifetime of a wireless sensor network. Recent techniques have shown the efficacy of multiple mobile elements (MMEs) in terms of energy consumption optimization. However, new challenges for energy-efficient MMEs scheme have emerged due to the emergence of large-scale WRSNs (LS-WRSNs) with big sensor-based data systems. Thus, large-scale deployments are currently limited owing to the bottlenecks of energy/lifetime, and mode of deployments of the sensor nodes. This article proposes a deadline-based MMEs (DB-MMEs) model exploiting the efficacies of the MMEs scheme to optimize energy consumption and provisioning. The DB-MMEs scheme exploits multifunctional wireless mobile charging vehicles (MCVs) for both wireless charging and data collection via a single-hop transmission. The scheme is specifically designed for delay-intolerant applications. None of the existing techniques has considered this approach to minimize latency and optimize energy consumption and provisioning for LS-WRSNs scenarios. The proposed scheme first organizes the sensors into several clusters for wireless charging and data collection. To optimize energy consumption and provisioning and address the challenges of energy/lifetime for LS-WRSNs scenarios, this article proposes analytical-based approaches to address some critical tradeoffs including: 1) determining the optimal amount of energy available for the MCVs; 2) finding the optimal number of MCVs deployed within a given deadline; and 3) finding the optimal number of data collection and charging points (DCCPs). Finally, the performance of the proposed approach is evaluated through experimental simulations, and the results validate the efficacy of the analytical-based method.

This article provides a survey about the state of the art in embedded intelligence (EI) research for smart cities. Currently, a comprehensive survey for EI research for smart cities is not available. This article presents a comprehensive review and discusses representative studies of the emerging and current paradigms for EI with the focus on the enabling technologies, applications, and challenges for smart cities from four areas: 1) first, the overview and classifications of the EI research are presented to show the full spectrum in this area, which also serves as a concise summary of the scope of this article; 2) second, the review and identification of interrelated enabling technologies in the form of EI platform technologies, and EI and device analytics technologies are discussed; 3) third, the article discusses various applications of EI utilizing these technologies and techniques for smart cities; and 4) the article also includes the challenges and insights for future research directions. This comprehensive survey article aims to give useful insights for the research area and motivate researchers toward the development of useful EI solutions for practical deployment in smart cities.

In machine learning, the nature of the dataset itself such as convexity of the data point sets affects the right choice of clustering algorithm to give good performance. This brief paper first focuses on how data convexity influences the clustering performance on biomedical datasets. Then it addresses the main challenges of two well-known clustering groups which are centroid-based and density-based clustering. These techniques typically require a set of parameters to be provided by the user before the algorithms can perform well in terms of good clustering and give the optimal number of clusters. Two parameter independent clustering techniques utilizing unique neighborhood sets (UNSs) called Parameter Independent Convex Centroid-based Clustering (ConvexClust) for convex-dominated datasets and Parameter Independent Non-Convex Density-based Clustering (NonConvexClust) for nonconvex-dominated datasets are introduced. The ConvexClust and NonConvex Clust algorithms are extensively evaluated on real-world biomedical datasets. Their performances are also compared with other clustering algorithms using evaluation criteria such as SSE, entropy and purity. The results have revealed the good performance of the proposed parameter-independent clustering techniques and also shown that most of the biomedical datasets in the experiments demonstrated their tendency towards convex-dominated data point sets.

It is a crucial need for a clustering technique to produce high-quality clusters from biomedical and gene expression datasets without requiring any user inputs. Therefore, in this paper we present a clustering technique called KUVClust that produces high-quality clusters when applied on biomedical and gene expression datasets without requiring any user inputs. The KUVClust algorithm uses three concepts namely multivariate kernel density estimation, unique closest neighborhood set and vein-based clustering. Although these concepts are known in the literature, KUVClust combines the concepts in a novel manner to achieve high-quality clustering results. The performance of KUVClust is compared with established clustering techniques on real-world biomedical and gene expression datasets. The comparisons were evaluated in terms of three criteria (purity, entropy, and sum of squared error (SSE)). Experimental results demonstrated the superiority of the proposed technique over the existing techniques for clustering both the low dimensional biomedical and high dimensional gene expressions datasets used in the experiments.

Affective computing is an emerging multidisciplinary research field that is increasingly drawing the attention of researchers and practitioners in various fields, including artificial intelligence, natural language processing, cognitive and social sciences. Research in affective computing includes areas such as sentiment, emotion, and opinion modelling. The internet is an excellent source of data required for sentiment analysis, such as customer reviews of products, social media, forums, blogs, etc. Most of these data, called big data, are unstructured and unorganized. Hence there is a strong demand for developing suitable data processing techniques to process these rich and valuable data to produce useful information. Early surveys on sentiment and emotion recognition in the literature have been limited to discussions using text, audio, and visual modalities. So far, to the author's knowledge, a comprehensive survey combining physiological modalities with these other modalities for affective computing has yet to be reported. The objective of this paper is to fill the gap in this surveyed area. The usage of physiological modalities for affective computing brings several benefits in that the signals can be used in different environmental conditions, more robust systems can be constructed in combination with other modalities, and it has increased anti-spoofing characteristics. The paper includes extensive reviews on different frameworks and categories for state-of-the-art techniques, critical analysis of their performances, and discussions of their applications, trends and future directions to serve as guidelines for readers towards this emerging research area.

This paper proposes an audio-visual emotion recognition system that uses a mixture of rule-based and machine learning techniques to improve the recognition efficacy in the audio and video paths. The visual path is designed using the Bi-directional Principal Component Analysis (BDPCA) and Least-Square Linear Discriminant Analysis (LSLDA) for dimensionality reduction and discrimination. The extracted visual features are passed into a newly designed Optimized Kernel-Laplacian Radial Basis Function (OKL-RBF) neural classifier. The audio path is designed using a combination of input prosodic features (pitch, log-energy, zero crossing rates and Teager energy operator) and spectral features (Mel-scale frequency cepstral coefficients). The extracted audio features are passed into an audio feature level fusion module that uses a set of rules to determine the most likely emotion contained in the audio signal. An audio visual fusion module fuses outputs from both paths. The performances of the proposed audio path, visual path, and the final system are evaluated on standard databases. Experiment results and comparisons reveal the good performance of the proposed system.

Due to the high levels of competition in a global market, companies have put more emphasis on building strong customer relationships and increasing customer satisfaction levels. With technological improvements in information and communication technologies, a highly anticipated key contributor to improve the customer experience and satisfaction in service episodes is through the application of video analytics, such as to evaluate the customer's emotions over the full service cycle. Currently, emotion recognition from video is a challenging research area. One of the most effective solutions to address this challenge is to utilize both the audio and visual components as two sources contained in the video data to make an overall assessment of the emotion. The combined use of audio and visual data sources presents additional challenges, such as determining the optimal data fusion technique prior to classification. In this paper, we propose an audio-visual emotion recognition system to detect the universal six emotions (happy, angry, sad, disgust, surprise, and fear) from video data. The detected customer emotions are then mapped and translated to give customer satisfaction scores. The proposed customer satisfaction video analytics system can operate over video conferencing or video chat. The effectiveness of our proposal is verified through numerical results.

Adaptive Educational Hypermedia (AEH) e-learning models aim to personalize educational content and learning resources based on the needs of an individual learner. The Adaptive Hypermedia Architecture (AHA) is a specific implementation of the AEH model that exploits the cognitive characteristics of learner feedback to adapt resources accordingly. However, beside cognitive feedback, the learning realm generally includes both the affective and emotional feedback of the learner, which is often neglected in the design of e-learning models. This article aims to explore the potential of utilizing affect or emotion recognition research in AEH models. The framework is referred to as Multiple Kernel Learning Decision Tree Weighted Kernel Alignment (MKLDT-WFA). The MKLDT-WFA has two merits over classical MKL. First, the WFA component only preserves the relevant kernel weights to reduce redundancy and improve the discrimination for emotion classes. Second, training via the decision tree reduces the misclassification issues associated with the SimpleMKL. The proposed work has been evaluated on different emotion datasets and the results confirm the good performances. Finally, the conceptual Emotion-based E-learning Model (EEM) with the proposed emotion recognition framework is proposed for future work.

Recent advances in large-scale networked sensor technologies and the explosive growth in big data computing have made it possible for new application deployments in smart cities ecosystems. In this paper, we define big sensor data systems, and survey progress made in the development and applications of big sensor data research. We classify the existing research based on their characteristics and smart city layer challenges. Next, we discuss several applications for big sensor data systems, and explore the potential of large-scale networked sensor technologies for smart cities in the big data era. We conclude this paper by discussing future work directions highlighting some futuristic applications. The aim of this survey paper is to be useful for researchers to get insights into this important area, and motivate the development of practical solutions toward deployment in smart cities.

Bottom-up visual saliency can be computed through information theoretic models but existing methods face significant computational challenges. Whilst nonparametric methods suffer from the curse of dimensionality problem and are computationally expensive, parametric approaches have the difficulty of determining the shape parameters of the distribution models. This paper makes two contributions to information theoretic based visual saliency models. First, we formulate visual saliency as center surround conditional entropy which gives a direct and intuitive interpretation of the center surround mechanism under the information theoretic framework. Second, and more importantly, we introduce a fast nonparametric multidimensional entropy estimation solution to make information theoretic-based saliency models computationally tractable and practicable in realtime applications. We present experimental results on publicly available eye-tracking image databases to demonstrate that the proposed method is competitive to state of the art.

In this paper, a region-based active contour model (ACM) with local information using watershed segmentation is proposed for lips contour detection. Compared to the ACM with global energy terms, the proposed system provides a more precise lips contour convergence under the circumstances where the lips are difficult to distinguish using global statistics. Furthermore, since the ACM is sensitive to the initial contour position, a modifiedH∞based on Lyapunov stability theory is proposed to provide better tracking of the subsequent lips feature points as the ACM initialization. The integration of the proposed ACM and modifiedH∞has revealed an improvement of the overall lips contour detection.

This paper presents a highly efficient audio-visual recognition system in compression domain. For face recognition systems, the multiband feature fusion method selects the wavelet subbands that are invariant to illumination and facial expression variations. These subbands will be extracted directly from the inverse quantization in the compression system. By taking the inverse quantized wavelet coefficient of the video as the input, the inverse wavelet transform which corresponds to image reconstruction is omitted. As a result, the computational complexity of the conventional video-based face recognition system is reduced. We also present a set of new face localization methods to localize the facial wavelet coefficients from the wavelet subband image. The dual optimal multiband feature fusion method is then used to fuse the two set of wavelet coefficients and generate the visual scores. Experimental results show that with low computational complexity, the proposed system achieves high recognition accuracy in UNMC-VIER, CUAVE, and XM2VTS audio-visual database.

Conventional face recognition suffers from problems such as extending the classifier for newly added people and learning updated information about the existing people. The way to address these problems is to retrain the system which will require expensive computational complexity. In this paper, a radial basis function (RBF) neural network with a new incremental learning method based on the regularized orthogonal least square (ROLS) algorithm is proposed for face recognition. It is designed to accommodate new information without retraining the initial network. In our proposed method, the selection of the regressors for the new data is done locally, hence avoiding the expensive reselecting process. In addition, it accumulates previous experience and learns updated new knowledge of the existing groups to increase the robustness of the system. The experimental results show that the proposed method gives higher average recognition accuracy compared to the conventional ROLS-algorithm-based RBF neural network with much lower computational complexity. Furthermore, the proposed method achieves higher recognition accuracy as compared to other incremental learning algorithms such as incremental principal component analysis and incremental linear discriminant analysis in face recognition.

This paper introduces a smoothing and preprocessing (S+P) technique for a line-based one-bit-transform (1BT) motion estimation scheme. In the proposed algorithm, a smoothing threshold ( Threshold S) is incorporated into the 1BT convolutional kernel. By using the smoothing threshold, scattering noise which is a common problem in most 1BT images can be greatly reduced. After the transformation, the 1BT images for the current and reference frames are divided into a number of macroblocks. The macroblock in the current frame is first compared with the macroblock at the same position in the reference frame. If the Sum of Absolute Difference (SAD) is below a certain preprocessing threshold ( Threshold P), the macroblock in the current frame is considered to have negligible movement and motion search is not performed. Simulation results show that this technique achieves high performance and greatly reduces the number of search operations. By incorporating the S+P technique, the PSNR achieved by the 1BT is approaches the performance of the 8-bit Full Search Block Matching Algorithm (FSBMA), and the difference is as low as 0.08 dB. In addition, this technique outperforms current state-of-the-art 1BT motion estimation techniques. An improvement in PSNR performance by up to 0.6 dB and a reduction in the number of search operations by 60% to 93% is achieved using video conferencing sequences.

A dynamic counterpropagation network based on the forward only counterpropagation network (CPN) is applied as the classifier for face detection. The network, called the dynamic supervised forward-propagation network (DSFPN) trains using a supervised algorithm that grows dynamically during training allowing subclasses in the training data to be learnt. The network is trained using a reduced dimensionality categorized wavelet coefficients of the image data. Experimental results obtained show that a 94% correct detection rate can be achieved with less than 6% false positives.

Images obtained with wavelet-based compression techniques such as set-partitioning in hierarchical trees (SPIHT) yield very good results. However, a lot of memory space is required as the wavelet coefficients for the whole image need to be stored for the process of set-partitioning coding. In this letter, we propose new virtual SPIHT tree structures for very low memory strip-based image compression. The advantage of the proposed work is that it reduces the memory requirements for practical software and hardware implementations significantly without sacrificing performance.