Real-Time Fire Detection: Integrating Lightweight Deep Learning Models on Drones with Edge Computing

Current indoor fire detection systems often have limitations, including insufficient intelligence and slow response times. Recent studies have started to explore AI vision-based methods for real-time fire and smoke detection to replace traditional fire sensors. However, indoor environments present many interference factors, and few methods can effectively demonstrate superior fire and smoke detection performance under multiple simulated interference conditions. To meet the requirements of real-time, high accuracy and anti-interference, a 3D convolutional neural network (3D-CNN)-based real-time fire and smoke detection model has been developed via red-green-blue (RGB) and near-infrared (NIR) feature fusion. RGB and NIR images with complementary information were used as dual-stream inputs to the model. The feature extraction module was enhanced by 3D attention mechanisms to simultaneously capture dynamic features between frames and static features within frames of fire and smoke. A non-linear feature gated fusion module has been developed to solve the challenges of missing multi-modal features and insufficient fusion. To validate the engineering effectiveness of the model, a large-scale multi-modal dataset was constructed in real indoor building scenarios, and a real-time indoor fire and smoke detection system was also developed and deployed at the experimental site.

Despite recent advances in deep learning, the rise of edge devices, and the exponential growth of Internet of Things (IoT) connected devices undermine the performance of deep learning models. It is clear that the future of computing is moving to edge devices. Autonomous vehicles and self-driving cars have leveraged the power of computer vision, especially object detection to navigate through traffic safely. Nevertheless, to be able to drive on all types of roads, these new vehicles have to be equipped with a road anomaly detection system, which strikes the need for small deep learning models of detecting road anomalies that can be deployed on these vehicles for safer driving experience. However, the current deep learning models are not practical on embedded devices due to the heavy resource requirements of the models, as such cannot be deployed on embedded devices. This paper proposes a theoretical approach to building a lightweight model from a cumbersome pothole detection model that is suitable on edge devices using knowledge distillation. It presents the theoretical approach of knowledge distillation, why it is a better technique of model compression compared to the rest. It shows that a cumbersome model can be made lightweight without sacrificing accuracy and with a reduced time complexity and faster training time.

Deep learning models have gained prominence in AI research due to their exceptional feature extraction capabilities. In fingerprint-based indoor localization, 1D and 2D convolutional neural networks (CNNs) are frequently utilized for their superior performance and minimal hardware requirements. However, these models primarily extract spatial features and often exhibit degraded real-time performance due to signal inconsistencies caused by shadowing, fading, and multipath effects. To address this, incorporating temporal variations through 3D CNNs can enhance outcomes, albeit with increased computational costs. This study proposes a knowledge distillation (KD) approach using 3D CNNs for indoor localization, to bridge the computational and performance gaps. The proposed KD framework employs a complex teacher model (TM) to train two simpler student models A (SM-A) and B (SM-B). We employ both response-based and feature-based KD methods to optimize the student models’ weights. The proposed techniques are evaluated using the UJILIB dataset collected from two floors of the University of Jaume I library. To enhance model performance, data augmentation is applied, yielding substantial improvements in test accuracy of 10.56% for TM, 16.19% for SM-A, and 16.93% for SM-B. The application of KD results in accuracy improvements of 5.45% for SM-A and 3.35% for SM-B, consequently reducing their average positioning errors (APEs) from 1.070 m and 1.030 m to 0.840 m and 0.944 m, respectively. Compared to KD-based SM-A, TM shows only 3.30% higher positioning accuracy, but with a 65.14% increase in computational complexity. Notably, the KD-enhanced SM-A also achieves a significantly lower APE of 0.84 m compared to the 2D CNN of 1.72 m.

Emotion recognition (ER) is crucial for understanding human behaviours, social interactions, and psychological well-being. Electroencephalography (EEG) has emerged as a promising tool for capturing the neural correlates of emotions. This work is a systematic review of articles in ER using EEG signals. A total of 120 articles from 1041 articles were selected based on PRISMA guidelines using defined inclusion and exclusion criteria, published between 2018 and 2024. This article aims to provide an in-depth understanding of the current landscape of ER from EEG signals utilizing deep learning (DL). This review offers valuable guidance for researchers and practitioners seeking more refined and reliable emotion classification systems. To explore the effectiveness of DL models in EEG-based ER, several potential DL models, such as convolutional neural network, long short-term memory (LSTM), gated recurrent unit (GRU), hybrid bidirectional LSTM (BiLSTM), bidirectional GRU, and advanced DL models such as convolutional recurrent neural network and EEG-Conformer models are applied to two popular datasets, SEED and GAMEEMO, respectively, to depict the full process of ER. Additionally, the performance of DL models is also compared with the performance of basic machine learning (ML) models such as SVM, k-nearest neighbors, logistic regression, and boosting algorithms such as AdaBoost, XGBoost and LightGBM. Through extensive experiments and performance evaluations, the performance of different models when applied to the datasets mentioned above is compared. The accuracy, precision, recall, and F1-scores are analysed to determine the most effective model for EEG-based ER. The findings of this study demonstrate that the performance of hybrid DL models is more efficacious than that of ML models. The best-performing model (BiLSTM) classified the emotions, with an accuracy of 90.54% when applied to the GAMEEMO dataset. This research contributes to the growing body of literature on ER and provides insights into the feasibility of using EEG signals to understand emotional states, and presents a structured roadmap for future exploration. The findings can aid in the development of more accurate and reliable ER systems, which can have wide-ranging applications in psychology, social sciences, and human-computer interactions.

Deep learning models have been shown to be vulnerable to adversarial attacks. Adversarial attacks are imperceptible perturbations added to an image such that the deep learning model misclassifies the image with a high confidence. Existing adversarial defenses validate their performance using only the classification accuracy. However, classification accuracy by itself is not a reliable metric to determine if the resulting image is "adversarial-free". This is a foundational problem for online image recognition applications where the ground-truth of the incoming image is not known and hence we cannot compute the accuracy of the classifier or validate if the image is "adversarial-free" or not. This paper proposes a novel privacy preserving framework for defending Black box classifiers from adversarial attacks using an ensemble of iterative adversarial image purifiers whose performance is continuously validated in a loop using Bayesian uncertainties. The proposed approach can convert a single-step black box adversarial defense into an iterative defense and proposes three novel privacy preserving Knowledge Distillation (KD) approaches that use prior meta-information from various datasets to mimic the performance of the Black box classifier. Additionally, this paper proves the existence of an optimal distribution for the purified images that can reach a theoretical lower bound, beyond which the image can no longer be purified. Experimental results on six public benchmark datasets namely: 1) Fashion-MNIST, 2) CIFAR-10, 3) GTSRB, 4) MIO-TCD, 5) Tiny-ImageNet, and 6) MS-Celeb show that the proposed approach can consistently detect adversarial examples and purify or reject them against a variety of adversarial attacks.

Deep neural networks (DNNs) are used in the environments with restricted resources. However, it is still generating many technical problems with memory, energy and computational constraints. This paper examines current developments in how to make deep learning models perform better with edge computing. The paper examines fundamental techniques such model trimming, quantization, knowledge distillation, and lightweight structure construction to simplify models while maintaining their correctness. Additionally discussed are ways to employ co-design methodologies and hardware-aware planning approaches to ensure model structures match the capabilities of edge hardware. The assessment also examines novel concepts such shared learning, TinyML, and edge-cloud teaming. According to the research, peripheral artificial intelligence must make trade-offs in several areas as it grows. It must strike a balance, for instance, between energy efficiency and working speed or between precision and delay. It also reveals how well systems of multi-objective planning handle these issues. Important elements influencing the path of edge intelligence are long-term application, real-time adaptability, and dynamic reasoning. Finally, the assessment highlights many issues still needing additional research including hardware separation, security flaws in networked systems, and the need of ongoing device learning. It calls for consistent toolchains and well-known standards that enable the responsible, flexible, and safe use of deep learning models on edge devices at last. Researchers and professionals should use this set of materials to create AI systems that are efficient, adaptable, privacy conscious, self-driving, real-world capable.

Maintaining wind turbine blades is a challenging task, often marked by high costs, safety risks, time inefficiency, and the possibility of incorrect diagnosis. A promising approach to support preventive maintenance involves the use of drones and deep learning for inspection and early fault detection. This paper offers a comparative study of five advanced deep learning models: Residual Networks (ResNet), Inception Networks (InceptionV3), YOLO (You Only Look Once), EfficientNet, and Vision Transformers (ViTs) applied to the DTU Drone Inspection Images of Wind Turbine dataset. The objective is to determine the most effective model for detecting defects in wind turbine blades, which is critical for ensuring the efficiency and safety of wind energy systems. The models were assessed using key performance metrics, such as accuracy, precision, recall, F1-score, and inference time. EfficientNet emerged as the top performer with the highest accuracy and balanced efficiency, while YOLO demonstrated the fastest inference time, making it ideal for real-time applications. ResNet and InceptionV3 also performed well, particularly in detecting subtle defects, whereas ViTs showed strong capabilities in capturing complex global features despite longer processing times. The findings provide valuable insights for selecting appropriate deep learning models in drone-based wind turbine inspections, contributing to more effective and reliable maintenance practices.

Microbiological analysis is crucial for identifying species, assessing infections, and diagnosing infectious diseases, thereby supporting both research studies and medical diagnosis. In response to these needs, accurate and efficient identification of bacterial colonies is essential. Conventionally, this process is performed through manual counting and visual inspection of colonies on agar plates. However, this approach is prone to several limitations arising from human error and external factors such as lighting conditions, surface reflections, and image resolution. To overcome these limitations, an automated bacterial colony counting and classification system was developed by integrating a custom-designed imaging device with advanced deep learning models. The imaging device incorporates controlled illumination, matte-coated surfaces, and a high-resolution camera to minimize reflections and external noise, thereby ensuring consistent and reliable image acquisition. Image-processing algorithms implemented in MATLAB were employed to detect bacterial colonies, remove background artifacts, and generate cropped colony images for subsequent classification. A dataset comprising nine bacterial species was compiled and systematically evaluated using five deep learning architectures: ResNet-18, ResNet-50, Inception V3, GoogLeNet, and the state-of-the-art EfficientNet-B0. Experimental results demonstrated high colony-counting accuracy, with a mean accuracy of 90.79% ± 5.25% compared to manual counting. The coefficient of determination (R2 = 0.9083) indicated a strong correlation between automated and manual counting results. For colony classification, EfficientNet-B0 achieved the best performance, with an accuracy of 99.78% and a macro-F1 score of 0.99, demonstrating strong capability in distinguishing morphologically distinct colonies such as Serratia marcescens. Compared with previous studies, this research provides a time-efficient and scalable solution that balances high accuracy with computational efficiency. Overall, the findings highlight the potential of combining optimized imaging systems with modern lightweight deep learning models to advance microbiological diagnostics and improve routine laboratory workflows.

In this work, we consider transferring the structure information from large networks to compact ones for dense prediction tasks in computer vision. Previous knowledge distillation strategies used for dense prediction tasks often directly borrow the distillation scheme for image classification and perform knowledge distillation for each pixel separately, leading to sub-optimal performance. Here we propose to distill structured knowledge from large networks to compact networks, taking into account the fact that dense prediction is a structured prediction problem. Specifically, we study two structured distillation schemes: i) pair-wise distillation that distills the pair-wise similarities by building a static graph; and ii) holistic distillation that uses adversarial training to distill holistic knowledge. The effectiveness of our knowledge distillation approaches is demonstrated by experiments on three dense prediction tasks: semantic segmentation, depth estimation and object detection. Code is available at https://git.io/StructKD.

Deep learning techniques have recently demonstrated remarkable success in numerous domains. Typically, the success of these deep learning models is measured in terms of performance metrics such as accuracy and mean average precision (mAP). Generally, a model's high performance is highly valued, but it frequently comes at the expense of substantial energy costs and carbon footprint emissions during the model building step. Massive emission of CO2 has a deleterious impact on life on earth in general and is a serious ethical concern that is largely ignored in deep learning research. In this article, we mainly focus on environmental costs and the means of mitigating carbon footprints in deep learning models, with a particular focus on models created using knowledge distillation (KD). Deep learning models typically contain a large number of parameters, resulting in a 'heavy' model. A heavy model scores high on performance metrics but is incompatible with mobile and edge computing devices. Model compression techniques such as knowledge distillation enable the creation of lightweight, deployable models for these low-resource devices. KD generates lighter models and typically performs with slightly less accuracy than the heavier teacher model (model accuracy by the teacher model on CIFAR 10, CIFAR 100, and TinyImageNet is 95.04%, 76.03%, and 63.39%; model accuracy by KD is 91.78%, 69.7%, and 60.49%). Although the distillation process makes models deployable on low-resource devices, they were found to consume an exorbitant amount of energy and have a substantial carbon footprint (15.8, 17.9, and 13.5 times more carbon compared to the corresponding teacher model). The enormous environmental cost is primarily attributable to the tuning of the hyperparameter, Temperature (τ). In this article, we propose measuring the environmental costs of deep learning work (in terms of GFLOPS in millions, energy consumption in kWh, and CO2 equivalent in grams). In order to create lightweight models with low environmental costs, we propose a straightforward yet effective method for selecting a hyperparameter (τ) using a stochastic approach for each training batch fed into the models. We applied knowledge distillation (including its data-free variant) to problems involving image classification and object detection. To evaluate the robustness of our method, we ran experiments on various datasets (CIFAR 10, CIFAR 100, Tiny ImageNet, and PASCAL VOC) and models (ResNet18, MobileNetV2, Wrn-40-2). Our novel approach reduces the environmental costs by a large margin by eliminating the requirement of expensive hyperparameter tuning without sacrificing performance. Empirical results on the CIFAR 10 dataset show that the stochastic technique achieves an accuracy of 91.67%, whereas tuning achieves an accuracy of 91.78%-however, the stochastic approach reduces the energy consumption and CO2 equivalent each by a factor of 19. Similar results have been obtained with CIFAR 100 and TinyImageNet dataset. This pattern is also observed in object detection classification on the PASCAL VOC dataset, where the tuning technique performs similarly to the stochastic technique, with a difference of 0.03% mAP favoring the stochastic technique while reducing the energy consumptions and CO2 emission each by a factor of 18.5.

In this paper, we proposed image processing technique for automatic real time fire and smoke detection in tunnel environment. To avoid the large scale of damage of fire occurred in the tunnel, it is necessary to have a system to minimize and to discover the incident as fast as possible. However it is impossible to keep the human observation of Closed-Circuit Television (CCTV) in tunnel for 24 hour. So if the fire and smoke detection system through image processing can warn fire state, it will be very convenient, and it can be possible to minimize damage even when people is not in front of monitor. The fire and smoke detection is different from the forest fire detection as there are elements such as car and tunnel lights and others that are different from the forest environment so that an indigenous algorithm has to be developed. The two algorithms proposed in this paper, are able to detect the exact position, at the earlier stage of incident. In addition, by comparing properties of each algorithm throughout experiment, we have proved the validity and efficiency of proposed algorithm

I. ABSTRACT Fire hazards pose a significant threat to life and property, making early detection crucial for minimizing damage. This project focuses on AI-based fire detection and alarming using computer vision and deep learning techniques. The system employs CCTV cameras or laptop cameras to capture real-time video feeds, which are analysed using Python, OpenCV, and deep learning models to detect fire accurately. A Convolutional Neural Network (CNN) or YOLO (You Only Look Once) object detection model is trained on fire datasets to distinguish fire from non-fire scenarios, reducing false alarms. Once fire is detected, the system triggers an alarm sound and can also send notifications via SMS, email, or IoT-based alerts to concerned authorities. This AI-powered solution is designed for high accuracy and efficiency, making it suitable for smart surveillance, industrial safety, and residential fire monitoring. The integration of edge computing devices like Raspberry Pi or Jetson Nano can further enhance real-time processing. The proposed system aims to improve fire safety measures by providing automated, real-time fire detection, reducing response time, and preventing large-scale disasters. II. KEYWORDS Fire Detection, Real-time Fire Monitoring, Computer Vision (CV).

Area under the curve may hide poor generalisation to external datasets

In daily life, fire accidents will happen at any time, but the way to put the fires off has not changed. Most of them are put out with water or chemicals, which cause secondary pollution to the surrounding environment and even cause irreversible damage to some items. The fire is very destructive, and people are paying more attention to fire prevention, so many places have prepared fire extinguishers. Public fire extinguishers include dry powder fire extinguishers, carbon dioxide fire extinguishers, foam fire extinguishers, but they will face some difficult problems in the application. Dry powder fire extinguishers and foam fire extinguishers will pollute the surrounding environment of the fire and they are not easy to restore to the previous state, beside, the wrong operation of carbon dioxide fire extinguishers in a confined space will also cause harm to the health of people in the fire scene. The new low-frequency acoustic fire extinguisher is proposed in this article. It could eliminate fire hazards and avoid secondary damage to the place. This type of fire extinguisher provides the least destructive and cleanest fire fighting method.

This study focuses on the field of pedestrian re-identification (ReID), aiming to enhance the accuracy and efficiency of individual recognition through advanced deep learning models. The article first introduces two deep learning models: Omni-Scale Feature Learning for Person Re-Identification (OSNet) and Multi-Scale Interaction Network (MSINet). OSNet employs Depthwise Separable Convolutions (DSC) and Omni-Scale Residual Blocks to improve the ability to learn features across different scales. MSINet, on the other hand, utilizes neural architecture search technology to design a lightweight network architecture, enhancing feature discrimination and flexible utilization through Twins Contrastive Mechanism (TCM) and Multi-Scale Interaction (MSI). Additionally, a Spatial Alignment Module (SAM) is proposed to enhance the consistency of images under different viewpoints or conditions. The experimental section selects two widely used pedestrian re-identification datasets, Market1501 and MSMT17, for evaluation, and the results show that MSINet outperforms existing methods in terms of accuracy and stability. The article concludes by summarizing the advantages of OSNet and MSINet in multi-scale feature learning and points out their application limitations.