Traditional evacuation path planning in fire scenarios typically depends on static building layouts and predefined escape routes, without accounting for dynamic threats such as fire spread and congestion. These factors may obstruct evacuation routes or expose evacuees to hazardous environments. Furthermore, the coupling between crowd dynamics and disaster evolution is often neglected, increasing the risk of secondary incidents such as congestion and stampedes. This paper presents an emergency evacuation framework that integrates Graph Neural Networks (GNNs) with Deep Reinforcement Learning (DRL) for multi-agent path planning. The proposed method employs GNNs to extract and aggregate environmental features from dynamic fire scenes, enhancing the perception capability of agents. DRL is then utilized to train agents for real-time obstacle avoidance and cooperative path optimization. Additionally, the strategy coordinates with electronic signage systems to adaptively guide evacuees. The effectiveness of the approach is validated through simulations under varying fire scenarios, where dynamic environment graphs are constructed using real-time temperature and crowd distribution data.

Deep reinforcement learning has shown remarkable capabilities in visual tasks, but it does not have a good generalization ability in the context of interference signals in the input images; this approach is therefore hard to be applied to trained agents in a new environment. To enable agents to distinguish between noise signals and important pixels in images, data augmentation techniques and the establishment of auxiliary networks are proven effective solutions. We introduce a novel algorithm, namely, saliency-extracted Q-value by augmentation (SEQA), which encourages the agent to explore unknown states more comprehensively and focus its attention on important information. Specifically, SEQA masks out interfering features and extracts salient features and then updates the mask decoder network with critic losses to encourage the agent to focus on important features and make correct decisions. We evaluate our algorithm on the DeepMind Control generalization benchmark (DMControl-GB), and the experimental results show that our algorithm greatly improves training efficiency and stability. Meanwhile, our algorithm is superior to state-of-the-art reinforcement learning methods in terms of sample efficiency and generalization in most DMControl-GB tasks.

This study aims to tackle the problems of fluctuation and overly high learning rates in the traditional Adabelief algorithm, together with its potential non-convergence during training. To mitigate these issues, we introduce Nadabmod, an adaptive algorithm with Nesterov acceleration and bounded momentum, to address learning rate escalation and enhance training stability and convergence. Firstly, an adaptive finite learning rate is introduced in the Adabelief algorithm framework, resulting in the Adabmod algorithm, which effectively resolves the issue of high learning rates during the training of neural networks, enhancing training stability. Subsequently, building on the classical first-order momentum term, the Nesterov momentum acceleration mechanism is integrated into the Adabmod algorithm to further improve efficiency. Through theoretical analysis of convergence in convex scenarios and experiments training diffusion models and generative adversarial networks, the efficacy of the Nadabmod algorithm is compared with that of Adam, Adabelief, Adamod, and Adabmod algorithms, confirming its superiority. Additionally, tests conducted with different initial learning rates and beta3 values verify the algorithm's robustness. Experimental results indicate that the Nadabmod algorithm significantly eliminates extreme learning rates and enhances training stability, effectively improving the training efficiency of neural network models.

Diffusion models have shown significant potential in generating high-quality time series data, especially in high-noise environments. However, the existing diffusion models primarily focused on extracting temporal features. Meanwhile, these models overlook the importance of modeling dynamic cross-channel relationships and sparse correlations in complex multivariate time series, which is insufficient for datasets with strong inter-channel dependencies. In this paper, we propose a clustering-enhanced conditional diffusion model (CE-Diff) that combines the advantages of adaptive channel clustering and conditional diffusion models. It maps time series data to the frequency domain to capture fine-grained correlations between channels. The clustering information is expressed as a sparse channel mask matrix, which dynamically guides the reverse denoising process through a masked attention mechanism. This clustering enhancement strengthens the diffusion model's ability to handle multivariate data. Experiments on several complex multivariate datasets have demonstrated that CE-Diff achieves higher accuracy and better generation quality in time series forecasting (TSF) compared to other generative models.

Image denoising has improved dramatically in re-cent years due to the rapid growth of neural networks. While some efforts have been made to train image denoising models using only single noisy images, current self-supervised denoising methods encounter challenges such as inefficient network training, loss of valuable information, or reliance on noise modeling. A random neighbor sub-sampler was introduced by Neighbor2Neighbor(NBR2NBR) [2] to create training image pairings, eliminating the requirement for noise-clean pairs or the necessity for multiple noisy observations. However, the training data obtained in this manner introduces a certain level of random noise during sub-sampling. Furthermore, the distribution of these noisy training pairs may not align well with the distribution of actual noisy image pairs, resulting in a data bias between the training and inference distributions. This data bias restricts the denoising performance of the method on real noisy images. This article proposes an effective method to reduce this data bias as much as possible. We introduce iterative data refinement, which trains the model on datasets with higher noise levels and refines noisy images for use in the next round using the learned model. This approach iteratively minimizes the bias between the generated noisier-noisier dataset and the ideal dataset to enhance the training of denoising models. Extensive experiments were carried out using synthetic and real datasets showcase that our approach surpasses previous works in performance.

Adaptive algorithms are extensively employed in the field of deep learning owing to their rapid convergence properties. Adam is the most common adaptive algorithm among them. However, it has revealed that Adam has a poor generalization ability. AdaBelief is an algorithm based on Adam with exact stepsize adjustment. It introduces a new second-order momentum that corrects Adam’s second-order moments. These algorithms still fail to converge during training due to the presence of instability and extreme learning rates. In this paper, we propose a new adaptive and momental bounded algorithm, called AdaBMod, which can effectively mitigate the sudden large learning rate problem and is especially suitable for training deep neural networks. By setting an adaptive finite learning rate in AdaBelief algorithm, the obtained AdaBMod can effectively eliminate the problem of high learning rate in the late training of neural networks, so as to make the training process more stable. Numerical simulation experiments on deep neural network tasks also show that our proposed AdaBMod algorithm eliminates the large learning rate during the training process. The results are also better than other current state-of-the-art optimizers.

The tail-l(1) minimization algorithm significantly enhances the recovery capability of sparse signals compared to the l(1) minimization algorithm. However, solving the tail-l(1) minimization problem requires high computational costs and a considerable amount of time. The normalized iterative hard thresholding (NIHT) algorithm, an improvement over the traditional IHT algorithm, exhibits good computational efficiency. Inspired by the NIHT algorithm, this paper introduces an enhanced NIHT algorithm, namely the tail-NIHT algorithm. The tail-NIHT algorithm retains the computational speed of the NIHT algorithm, greatly improving the efficiency of solving the tail-l(1) minimization problem. Additionally, the tail-NIHT algorithm enhances the sparse signal recovery capability of the NIHT algorithm. Experimental results demonstrate that this algorithm is a promising approach in compressed sensing.

The application of medical imaging holds significant value in the identification and monitoring of neurological conditions. The visualization of brain MRI data assists medical practitioners in understanding complex brain structures and identifying irregularities. Transfer learning augments the efficacy of convolutional neural networks (CNNs), which are renowned for their exceptional capability in medical image interpretation. Utilizing UNet and FPN neural networks in tandem with transfer learning techniques, the current study seeks to evaluate the precision of brain MRI data visualization. The prevalent UNet and FPN architectures for medical image analysis are described in this study. The UNet model employs skip connections in its encoder-decoder architecture, while the FPN model collects multi-scale data using a feature pyramid. The effectiveness of segmentation and classification schemes has been demonstrated. To evaluate transfer learning, UNet, and FPN networks are trained to recognize generic image representations using a massive dataset like ImageNet. Brain MRI imaging is employed as target data to optimize the networks. Pre-training networks are intended to be equipped with sophisticated and comprehensive source domain features, with the specific purpose of facilitating brain MRI visualization. The information transmission procedure improves the ability of networks to discern nuanced attributes and faithfully depict them. The novel model framework implemented in this study is predicated on a histogram-based threshold. Through the examination of the image intensity histogram, we successfully computed the suitable threshold and evaluated the model’s proficiency in segmenting brain lesions in magnetic resonance imaging (MRI) scans. Utilizing the proposed method, correlations between numerous modalities were exploited, as each modality produced differing degrees of accuracy, which peaked at 91%. A combination of bottom-up and top-down approaches was utilized to integrate the numerous modalities into the model. By effectively capitalizing on the complex interrelation among modalities, our methodology improves the precision of segmentation. In the domain of brain MRI data visualization, the findings emphasize the importance of implementing transfer learning methodologies for medical image processing. By employing pre-trained networks, medical practitioners can more precisely evaluate brain architecture and diseases. This article makes a scholarly contribution to the expanding field of study that explores the feasibility of utilizing transfer learning to improve the performance of Convolutional Neural Networks (CNNs) in the context of medical imaging.

In the field of image processing, total variational model is an effective prior model. In order to better eliminate impulse noise, an effective method is to use ℓ _1-norm total variational model. However, the TV image recovery always produces staircase artifacts, and the ℓ _1-norm excessively penalizes the signal entries. Therefore, in this paper, a new total variational model is proposed to eliminate the staircase effects and impulse noise. We use ℓ _0-norm as the data fidelity term to eliminate impulse noise and the hybrid total variation as the regularization term to effectively eliminate staircase artifacts. In order to effectively tackle the proposed ℓ _0-norm and hybrid total variation, we first express this problem as a Mathematical Program with Equilibrium Constraints(MPEC), and then a Proximal Alternating Direction Method of Multipliers(PADMM) is adopted to solve this problem. In the experimental part, we adopt three indices: SNR_0 , SNR_1 and SNR_2 to measure the quality of image restoration. When the value of the SNR tends to be stable, then each algorithm stops iteration. Numerical simulation results show that the proposed method has better performance in removing impulse noise, suppressing staircase effect and preserving image edge information.

In recent years, deep learning has dramatically improved state of the art in many practical applications. However, this utility is highly dependent on fine-tuning of hyperparameters, including learning rate, batch size, and network initialization. Although many first-order adaptive gradient algorithms (e.g., Adam, AdaGrad) have been proposed to adjust the learning rate, they are vulnerable to the initial learning rate and network structure in the training over-parameterized models, especially in the dynamic online setting. Therefore, the main challenge of using deep learning in practice is how to reduce the cost of tuning hyperparameters. To address this problem, we integrate the adaptive strategy of Radhakrishnan et al. and the acceleration strategy of Ghadimi et al. to propose a fast adaptive online gradient algorithm, FAOGD. The adaptive strategy we adopt only adjusts the learning rate according to the historical gradient and training loss value, while the acceleration strategy is the heavy-ball momentum used to accelerate the training of deep models. The proposed FAOGD enjoys merit that there is no need to tune hyperparameters related to the learning rate, which thus saves much unnecessary computational overhead. It is also shown that FAOGD can obtain the regret bound of O( √(T)) , matching the Adam and AdaGrad using the empirical learning rate. Simulation results in the over-parameterized neural networks clearly show that FAOGD outperforms existing algorithms. Furthermore, FAOGD is also robust to network structures and batch size.

The Progressive Mean (PM) control chart is a widely recognized tool to notice the insignificant and standard variations in the process location parameter. There is one deficiency in PM chart, it generates signals which are out of control, and when the standard deviation is processed this deficiency changes the results. To overcome this problem, we proposed a method in case of not stable process stand deviation chart is used, which enables monitoring of process which is more robust in this case. The suggested chart is a participant for and charts. The numerical results concluded that the performance of the proposed chart is superior to detect small-scale and standard changes in the process parameter. To support the study an expressive application is also provided.

The key to improving the fusion quality of infrared–visible images is effectively extracting and fusing complementary information such as bright–dark information and saliency details. For this purpose, an improved hybrid multiscale fusion algorithm inspired by non-subsampled shearlet transform (NSST) is proposed. In this algorithm, firstly, the support value transform (SVT) is used instead of the non-subsampled pyramid as the frequency separator to decompose an image into a set of high-frequency support value images and one low-frequency approximate background. These support value images mainly contain the saliency details from the source image. And then, the shearlet transform of NSST is retained to further extract the saliency edges from these support value images. Secondly, to extract the bright–dark details from the low-frequency approximate background, a morphological multiscale top–bottom hat decomposition is constructed. Finally, the extracted information is combined by different rules and the fused image is reconstructed by the corresponding inverse transforms. Experimental results have shown the proposed algorithm has obvious advantages in retaining saliency details and improving image contrast over those state-of-the-art algorithms.

Adaptive algorithms are widely used in deep learning because of their fast convergence. Among them, Adam is the most widely used algorithm. However, studies have shown that Adam’s generalization ability is weak. AdaX is a variant of Adam, which introduces a novel second-order momentum, modifies the second-order moment of Adam, and has good generalization ability. However, these algorithms may fail to converge due to instability and extreme learning rates during training. In this paper, we propose a new adaptive and momental bound algorithm, called AdaXod, which characterizes of exponentially averaging the learning rate and is particularly useful for training deep neural networks. By setting an adaptively limited learning rate in the AdaX algorithm, the resultant AdaXod can effectively eliminate the problem of excessive learning rate in the later stage of neural networks training and thus results in stable training. We conduct extensive experiments on different datasets and verify the advantages of the AdaXod algorithm by comparing with other advanced adaptive optimization algorithms. AdaXod eliminates large learning rates during neural networks training and outperforms other optimizers, especially for some neural networks with complex structures, such as DenseNet.

Bright-dark components and edge details are the most important complementary information between infrared and visible images. To extract and fuse them efficiently, a novel non-subsampled morphological fusion algorithm is proposed in this paper. The algorithm uses non-subsampled pyramid ( NSP ) as the spatial-frequency splitter to decompose the source image to get a series of high-frequency detail images and one low-frequency background image. Then, a dual-channel multi-scale top–bottom hat ( MTBH ) decomposition is constructed to extract the bright-dark details from the low-frequency background. In addition, to extract the edge details with different directions from high-frequency images, a dual-channel multidirectional inner-outer edge (MIOE) decomposition is constructed. Through these decompositions, the bright-dark information and edge details present in the source images can be effectively extracted. Then, based on the distinct roles of the extracted information, the decomposed images are fused using diverse fusion strategies. Subsequently, the fused image is reconstructed using the appropriate inverse transforms corresponding to each decomposition. The experimental results demonstrate that the fusion images generated by this algorithm exhibit richer details and higher image contrast compared to those produced by state-of-the-art algorithms.

Adaptive algorithms are widely used in deep learning because of their fast convergence. Among them, Adam is the most widely used algorithm. However, studies have shown that Adam's generalization ability is weak. AdaX is a variant of Adam, which modifies the second moment of Adam and has good generalization ability. We propose a new adaptive and momental bound algorithm, called AdaXod, which characterizes of exponentially averaging the learning rate and is particularly useful for training deep neural networks. By setting an adaptively limited learning rate in the AdaX algorithm, the resultant AdaXod can effectively eliminate the problem of excessive learning rate in the later stage of neural network model training and thus stabilize training. Simulation experiments verify that AdaXod eliminates large learning rates during neural network training and outperforms other optimizers, especially on the complex network structures such as DenseNet.

A tail-Hadamard product parametrization (tail-HPP) approach is proposed for sparse signal recovery in compressed sensing. The algorithm has both the efficiency of the HPP technique and the much greater capacity of signal recovery enabled by the tail- t 1 -minimization approach. We prove that the tail-HPP approach is equivalent to the tail- t 1 -minimization problem. The efficiency of the tail-HPP algorithm is clearly evident compared to direct solution approaches of the tail- t 1 -minimization problem. These superiority of the tail-HPP algorithm is confirmed by extensive simulation experiments in comparison with state-of-the-art sparse recovery techniques.

Image restoration is a serious inverse problem, and the regularization method is recognized as an effective method. In this paper, a ℓ 0 non-convex hybrid total variation(NHTV) regularization method is proposed, and the TV-based restoration problem is solved by using ℓ 0 norm data fidelity. In order to effectively tackle the proposed non-convex and non-smooth optimization problem, we first express this problem as a Mathematical Program with Equilibrium Constraints(MPEC), and then a proximal Alternating Direction Method of Multipliers(PADMM) is adopted to solve this problem. Numerical simulations verify the effectiveness of our proposed method. The comparations with other convex TV-based regularization methods are also conducted, which clearly show that this method can achieve better suppression of the staircase effect, effectively preserve the edge information and therefore obtain a higher signal-to-noise ratio.

Since the breakout of Corona Virus Disease 2019 (COVID-19), the global fight against influenza has begun. Var-ious technologies have been developed to support the fast-growing contactless service market, and hence contactless services are rapidly becoming a new growth strategy. In particular, the retail service industry most urgently needs contactless service technology. A representative technical case is the self-checkout machine, which can reduce labor costs and provide customer satisfaction. We present a solution in this article. We propose a hand gesture recognition contactless self-checkout system, which is a hand gesture recognition model based on YOLOv5s. The hand gesture recognition mAP (0.5) value reaches 0.995, the mAP (0.5:0.95) value reaches 0.865, and the Fl score is 0.96, together with the accuracy and recall rate is close to 1. Compared with the excellent algorithm YOLOx-s, the FPS value of YOLOv5s can reach 123 (YOLOx-s is 108). In addition, the model can be used to detect recorded static and dynamic hand gestures in real-time. Practical results show that the YOLOv5s can effectively recognize hand gestures and realize the contactless checkout process.

This paper investigates an adaptive gradient‐based online convex optimization problem over decentralized networks. The nodes of a network aim to track the minimizer of a global time‐varying convex function, and the communication pattern among nodes is captured as a connected undirected graph. To tackle such optimization problems in a collaborative and distributed manner, a weight decay distributed adaptive online gradient algorithm, called WDDAOG, is firstly proposed, which incorporates distributed optimization methods with adaptive strategies. Then, our theoretical analysis clearly illustrates the difference between weight decay and L2 regularization for distributed adaptive gradient algorithms. The dynamic regret bound for the proposed algorithm is further analyzed. It is shown that the dynamic regret bound for convex functions grows with order of O(n(1+logT)+nT) , where T and n represent the time horizon and the number of nodes associated with the network, respectively. Numerical experiments demonstrate that WDDAOG works well in practice and compares favorably to existing distributed online optimization schemes.

This paper investigates an online convex optimization problem on time-varying directed networks, where each agent holds its own convex cost function and the goal is to cooperatively minimize the sum of the global cost function. To tackle such optimization problems, an accelerated distributed online gradient push-sum algorithm is firstly proposed, which combines the momentum acceleration technique and push-sum strategy. Then, we specifically analyze the regret for the proposed algorithm. The theoretical result shows that the individual regret of the proposed algorithm achieves an improved rate with order of $\mathcal{O}\left( {\sqrt {1 + \log T} } \right)$, where T is the time horizon. Moreover, we implement the proposed algorithm in sensor networks for solving the distributed online estimation problem, and the results demonstrate the effectiveness of the proposed algorithm.