Machine learning (ML) techniques are applied for profiling computing and processing resources data collected while running deep neural network models on edge devices. Adaptive deep neural network (DNN) model switching requires proper benchmarking for categorizing AI models based on their applications and computational resources enabled by their processing accelerators. Based on benchmark metrics, DNN models can be classified into tiny, low, small, medium, and large resources, then identify DNN models that perform well within resource constraints. Ensure efficient resource allocation, latency management, and trade-off between accuracy and resource. In this work, we propose a benchmark for edge transfer artificial intelligence learning service (TALS) that uses ML techniques. They aim at classifying DNN models by their target edge applications while running edge inferences. We used both unsupervised learning (UL) and supervised learning (SL) techniques to identify the most effective features for the TALS models and to benchmark the performance of edge devices. To achieve this, two approaches were investigated: first, determining features based on edge inference’s computing resources profiling using principal component analysis; and second, classifying the DNN models at the target application level using a regression approach based on historical resource utilization data. In addition, we propose a dynamic model transfer learning that switches between a set of pre-trained and optimized and quantized DNN models based on the cost function. ML techniques learn resource-aware prediction from new resource allocation data and ensure that the multicriteria switch cost selects the inference task models that meet the edge resource constraint requirements. The experimental results highlight a strong relationship between the supervised learning model and the clustering execution method. The dynamic switching approach on real edge devices demonstrates dynamic switching between models according to inference task complexity. We conclude that dynamic switching models allows to ensure smooth operation without overloading resources in edge intelligence.

Infrared thermal imaging presents significant potential in various domains. However, the widespread development of this technology is hindered by the high cost associated with acquiring high-quality thermal imaging sensors. To overcome this challenge, super-resolution techniques have emerged as a viable solution for extracting valuable information from low-resolution thermal images. While generative adversarial networks (GANs) have been widely adopted for thermal imaging super-resolution, their performance is limited by the inherent lack of detail in low-resolution training images, resulting in reduced fidelity and accuracy in generating high-resolution reconstructions. To tackle this challenge, this letter introduces FEGAN, a novel approach that enhances the performance of GANs by incorporating a feature-oriented enhanced (FE) mechanism within the generative network (GN). The FE plays a pivotal role in extracting high-frequency texture and edge details from lowresolution inputs and reconstructing them into enhanced images. This process substantially improves textures and edges within the training set of thermal images. Furthermore, refinements have been applied to both the GN and the discriminative network (DN) to enhance feature extraction efficiency. The experimental findings unequivocally demonstrate the superior performance of FEGAN compared to state-of-the-art methods. FEGAN achieves impressive performance metrics, including PSNR of 27.18, SSIM of 0.6523, FSIM of 0.5500, and LPIPS of 0.1221, highlighting its remarkable capabilities in the realm of thermal image superresolution.

Indoor positioning in the firefighting ground shows promising application prospects for enhancing rescue safety and efficiency. Low visibility and signal interference caused by smoke pose significant difficulties for visual Simultaneous Localization and Mapping (SLAM) systems and radio frequency-based localization methods, while the performance of existing pedestrian dead reckoning (PDR) methods is affected by unpredictable user gaits. This paper introduces a gait analysis-based PDR (GA-PDR) for deriving heading estimation in PDR to improve its localization performance. The proposed method determines the step pattern by analyzing the features of inertial measurement unit data, thereby enabling the classification of forward, left- and right-turn and around-turn from left or right-side movements. In addition, this study introduces a redundant turn elimination method to differentiate false positive patterns via a time-domain heading analysis for turn movements. The location tracking performance with the proposed GA-PDR approach is validated using a self-created dataset established under a smoke-filled experiment, the results of which indicate a lower loop closure error compared with the traditional PDR in all of the tested scenarios.

Edge intelligence (EI) refers to a set of connected systems and devices for artificial intelligence (AI) data collected and learned near the data collection site. The EI model inference phase has been improved through edge caching technologies such as intelligent models (IMs). IM inference across heterogeneously distributed edge nodes is worthy of discussion. The present focuses on software-defined infrastructure (SDI) and introduces a containerized EI framework for a mobile wearable Internet-of-Things (IoT) system. This framework, called the containerized edge intelligence framework (CEIF), is an inter-working architecture that allows the provisioning of containerized EI processing intelligent services related to mobile wearable IoT systems. CEIF enables dynamic instantiation of the inference services of AI models that have been pre-trained on clouds. It also accommodates edge computing devices (ECDs) running the container virtualization technique. Dynamic AI learning policies can also help with workload optimization, thereby reducing the response time of the requests of the EI inference. To stall the rapid increase in user workload when inferring the collected data for analysis, we then propose a deep q-learning algorithm in which the container cluster platform learns the varying user workload at the location of each ECD. The requests of the EI inference are scaled with the learned value and are processed successfully without overloading the ECD. When evaluated in a case study, the proposed algorithm enabled scaling of the processing requests of the EI inference in a containerized EI system while minimizing the number of instantiated container EI instances. The EI inference's requests are completed in an under-loaded container EI cluster system.

With the advancement of Internet-of-Things (IoT) technology, various domains, such as smart factories and smart cars have used this new technology to provide value-added services. In addition, technologies, such as multiaccess edge computing (MEC) and network slicing provide another opportunity for the IoT to support more advanced and real-time services that could not have been previously supported. However, the simple integration of such technologies into the IoT does not take the full advantage of MEC and network slicing or the reduction of latency and traffic prioritization, respectively. Therefore, there is a strong need for an efficient integration mechanism for IoT platforms to maximize the benefit of using such technologies. In this article, we introduce a novel architectural framework that enables the virtualization of an IoT platform with minimum functions to support specific IoT services and host the instance in an edge node, close to the end user. As the instance provides its service at the edge node, where the MEC node and network slice are located, the traffic for the end user does not need to traverse back to the cloud. This architecture guarantees not only low latency but also efficient management of IoT services at the edge node. To show the feasibility of the proposed architecture, we conducted an experimental evaluation by comparing the transmission time of both IoT services running on the central cloud and those using sliced IoT functions in the edge gateway. The results show that the proposed architecture provides twice as much transmission speed as that from the conventional cloud-based IoT platform.

With the advancement of the Internet of Things (IoT), different use cases, such as smart factories, smart cities, and smart cars are being developed. Technologies, such as multiaccess edge computing (MEC) and network slicing are being researched to support various vertical use cases. However, simply utilizing network slicing and MEC is insufficient when providing a specific IoT service. Accordingly, there is a need for a method of arranging and operating common service functions (CSFs) used in the IoT platform at the edge of the network. In addition, it is necessary not only to deliver virtual CSFs but also to implement autoscaling policies of these virtual common service resources. This study extends the idea of resource slicing technologies for the IoT service layer. The IoT service layer exposes the virtual IoT CSFs as an IoT slice service to be provided in the form of virtual network functions (VNFs) with MEC applications on top of the network function virtualization infrastructure (NFVI) at the edge of the network. We propose a framework architecture for virtual IoT slice service orchestration, and illustrate how to instantiate the virtual IoT service functions in the existing centralized cloud to the MEC platform. We also propose an elastic computing algorithm of virtual IoT slice services functions(vIoT-SSFs) resources at the NFVI so that the IoT applications and underlying IoT resources can access vIoT-SSFs at the edge of the network.