Facial video-based remote physiological measurement (rPPG) can non-invasively estimate vital signs, such as heart rate (HR), which often faces challenges under varying lighting conditions. We propose the PhysLight framework to enhance the accuracy of rPPG heart rate measurement through adaptive video relighting. Our approach subtly modifies illumination in video frames to improve detection accuracy while maintaining visual quality. The framework includes a GenLightNet to extract ideal lighting priors and a WipeLightNet module to refine poorly lit videos. Extensive evaluations on benchmark datasets show that our method significantly improves HR estimation reliability, outperforming existing baselines and enhancing non-contact physiological monitoring in diverse environments.

Large language models (LLMs) provide essential capabilities for smart systems, yet navigating complex reasoning frontiers reliably remains challenging, hindering deployment in dynamic environments. Existing prompting methods often lack robustness or demand costly multi-step interaction. We introduce Twin Prompt, a novel automated framework inspired by human cognition, operationalizing structured problem reformulation and answer refinement within a single, end-to-end interaction requiring no manual examples. This cognitively grounded structure guides the LLM’s internal reasoning, enhancing analysis and leveraging latent self-correction capabilities to improve accuracy and reliability. Evaluations on challenging mathematical and general reasoning benchmarks (GSM8K, MATH, MMLU, BBH) demonstrate Twin Prompt significantly boosts performance over standard baselines across diverse LLMs. These findings highlight the potential of structured, single-pass prompting to advance LLM reasoning, enabling more capable and dependable AI components for navigating a dynamic world.

In traditional office management, it is labor-intensive to perform real-time oversight on equipment and personnel. To address this challenge, this paper proposes a digital twin-based office management system. The system leverages the ESP8266 wireless module for device control and data collection, and employs the YOLOv5 deep learning model for real-time detection of employees’ working conditions. Additionally, a virtual office environment is constructed using the Unity engine. The system implemented herein enables real-time monitoring and analysis of office utilization, and assists managers in optimally allocating resources to enhance resource utilization efficiency by leveraging intelligent sensing and decision-making technologies. The system incurs low hardware and software costs, minimal data transmission latency, and rapid response times across its modules. Moreover, through data masking techniques, the system can protect the privacy of office personnel while enabling real-time monitoring.

Video Salient Object Detection (VSOD) is a significant pre-work for many vision applications. Different for Salient Object Detection (SOD), an effective VSOD model requires not only the spatial domain of origin image but also temporal domain. In this paper, we proposed an optical flow guided pyramid network (OFPN) for VSOD, which exploit the temporal optical flow (OF) to assist VSOD. Due to the fact that optical flow maps have slightly lower quality compared to depth maps, we designed two modules for seeking better improvement. To this end, we render optical flow maps from RGB images firstly. Then, an adaptive cross-modal attention module (ACA) is designed for multi-modal fusion. The high-level encoded features are aggregated into a shared decoder for primary prediction. Besides, the low-level features are separately sent into multi-scale context attention module (MCA) for multi-scale context fusion with the assist of the primary prediction level by level. Further, we exploit a multi-scale loss to take full advantage of the hierarchical details through image pyramid structure. Extensive experiments on five benchmark datasets demonstrate the superiority of our method against 12 state-of-the-art methods.

This paper presents a Unity engine-based digital twin intelligent decision-making system for campuses, aiming to improve campus management efficiency and student experience. The system combines shapefile information and tilt-shot fusion technology to achieve 3D reconstruction of campus buildings and environments, creating a digital twin model. For intelligent decision-making, we simulated a virtual energy environment and applied a reinforcement learning algorithm to address real-world energy decision challenges. Concurrently, IoT technology is used to monitor campus devices and resources, including energy utilization, security, and environmental quality. The integrated data is presented through a visual and interactive interface for real-time monitoring and management of campus resources by administrators and students. This system enhances resource utilization efficiency, sustainability, and security, showcasing the practical application of digital twin technology in modernizing campus management and improving the student experience in institutions.

In the fields of autonomous driving and robotics, text image super-resolution can improve the resolution of the image obtained by the device, and help the system capture text details and long-distance text in the scene to improve the perception and decision-making ability. Significant progress has been made in prior-based text image super-resolution in recent years. However, the fusion of complex languages text prior information, such as Chinese, leads to a sharp increase in network parameters and computational costs, limiting the application of resource-constrained mobile robots in corresponding scenarios. In response to this issue, we propose a lightweight scene text image super-resolution method that incorporates semantic priors (MPT-TISR). Firstly, image convolutional layer and text recognizer are employed to extract low-resolution image features and text sequences, respectively. Then, an efficient Text Semantic Feature Fusion Block (FPCAT) based on Transformers and Principal Component Analysis (PCA) is constructed to associate essential text information with image features. Simultaneously, an improved MobileViTv3-based Sequential Residual Block (SRBMVT3+) is designed to learn high-dimensional deep features and enhance feature representation capabilities. Additionally, a binary gradient loss function is utilized to filter noise and guide the reconstruction process towards focusing on text details. Experimental results on a Chinese scene dataset demonstrate that MPT-TISR outperforms existing methods in terms of reconstructed image quality and significantly improves the recognition accuracy in the downstream text recognition task.

Chain of thought (CoT) prompting drives complex reasoning in large language models (LLMs), but remains scarcely explored for smaller Base Language Models (BLMs). We pioneer the Heuristic Chain of Thought (HCoT) approach for BLMs simply via "Let’s use knowledge" prompts. Further, we bridge the lack of guidance in HCoT by innovating SPIRE, a template providing specificity, purpose, information, role & capacity and expression to efficiently apply knowledge. Experiments show that combining HCoT with the SPIRE format significantly improves BLMs performance on question answering and translation tasks after minimal tuning. For example, on SQuADv1.1, our method increases the character F1 by 4.12% over zero-shot CoT using a 41.7M parameter BERT model.

There is an urgent need for industrial manufacturing to fully integrate with emerging technologies to build enterprise core competitiveness. Currently, existing methods have difficulty meeting the high-precision and stability practical requirements with diversified industrial products. In this study, a PMAE (Parallel Multiple AutoEncoders, PMAE) model is proposed, which is designed with parallel multiple encoders based on the AutoEncoder framework. It uses the network of parallel multiple encoders with different encoder structures to obtain latent features that have rich and precise semantic information. A consistency objective function is proposed to make the PMAE network converge stably and rapidly, which allows the aggregated latent features to be simultaneously reconstructed and adaptively classified. Compared with the state-of-the-art methods on the NEU-CLS, Data-Crack, and DAGM2007 datasets, our method achieves the most stable performance and the highest accuracy in different defect detection tasks.

The need for cancelable biometric techniques has seen a progressive rise due to the rapid deployment of biometric authentication systems. These techniques prevent compromising biometric data by generating and using their corresponding cancelable templates for user authentication. However, the non-invertible distance preserving transformation methods employed in various schemes are often vulnerable to information leakage since matching is performed in the transform domain. This paper proposed a non-invertible distance preserving scheme based on vector permutation and shift-order process. First, the dimension of feature vectors is reduced using kernelized principal component analysis before randomly permuting the extracted vector features. A shift-order process is then applied to the generated features to achieve non-invertibility and combat similarity correlation-based attacks. The generated hash codes are resilient to various security and privacy attacks such as ARM, masquerade, and brute-force preimage. Experimental evaluations conducted on eight fingerprint datasets from FVC2002, FVC2004, and FVC2006 reveal a high matching performance of the proposed method with better recognition accuracy than other existing state-of-the-art. The scheme also fulfills the revocability and unlinkability requirements of cancelable biometrics.

In high-voltage power systems, insulators are essential components in transmission lines for increasing shooting distance and securing wires. Unmanned aerial vehicle imaging becomes a common way of inspecting the state of the insulators. However, the automatic detection of insulators with complex backgrounds is still a challenging task. Most of the existing object detection methods are based on anchors, which do not have sufficient ability to describe objects that have a string-like structure. To tackle it, inspired by the keypoints-based object detection method, we propose a novel chain structure framework to detect insulators. First, we model the string-like object as a chain structure consisting of keypoints and their linkage, by which the insulator strings features are efficiently encoded and trained in the proposed ChainNet. Then, an assembling algorithm is proposed to assemble the estimated keypoints and linkages into chains, by which the insulator strings can be tightly enclosed in rotational bounding boxes. We evaluate the proposed approach on our collected large-scale rigorous dataset under the Directed Intersection over Union metric. The extensive experimental results show that the proposed method achieves 75.7% mAP, which yields 3.4, 15, and 10.6 improvements to the state-of-the-art rotational anchor, axial-aligned anchor, and anchor-free detection methods, respectively. Moreover, the proposed framework can easily be extended to detect other string-like manmade objects in the industrial area.

Due to their powerful modeling and reasoning capabilities, graph neural networks have not only achieved adequate performance in unstructured data but, in recent years, their research interests in Euclidean data such as images have also been on the rise. In this context, the most common task is image classification, whose method, based on graph neural networks, is roughly divided into two stages. The first is the graph construction stage, where the images are converted into graph structure data (composed of a node-edge-node form). The second is the graph classification stage, where the processed graph data is loaded into the graph classification network for graph classification. Subsequently, the images classification results are obtained. However, nearly two problems are faced in this setting. One is that the graph construction stage takes a long time due to the considerable computation that is required in order to convert images into graph structured data, the other problem is related to the graph classification used in the graph classification stage. The number of layers in the network tends to be limited, usually 4 layers or less. Hence, the graph’s classification accuracy is affected to some extent. Subsequently, this study proposes a deep graph neural image classification model based on gSLIC, combining the attention mechanism to conduct related experiments, in order to prove that while the speed of graph construction is greatly improved, the image classification accuracy exceeds that of most existing models based on graph neural networks for image classification.

Anomaly detection technology relying on memory reconstruction leverages the difference in reconstruction errors between the normal and abnormal frames to achieve superior detection performance. However, there are still some challenges with this technology. First, the memory has insufficient representation capacity for features. Second, there is a contradiction between feature fusion and reconstruction. As feature fusion copies the abnormal patterns into the reconstructed frames, the abnormal frames are effectively reconstructed, reducing the detection performance. In response to these challenges, we use a memory update threshold to improve the representational power of memory. We also propose a dual-encoder anomaly detection model to restrict anomaly feature propagation. Experiment results demonstrate the effectiveness and robustness of our approach.

Abstract Small objects only occupy a few pixels in an image, which results in low performance of small object detection for existing object detection algorithms. Therefore, the authors propose a dense sampling and detail enhancement network (DSDE‐Net) to address this issue. The network contains a dense sampling module used to increase the resolution of feature maps and expand the receptive field, which includes an atrous spatial pyramid pooling network and a coordinate attention mechanism to systematically process feature maps. Simultaneously, the authors introduce a detail enhancement branch that contains edge and detailed information to generate detailed enhancement feature maps through Gaussian filtering to compensate for the loss of small object information that occurs in the feature extraction process. The experimental results demonstrate that the proposed network outperformed related methods. Compared with the state‐of‐the‐art algorithm DetectoRS, it effectively achieves approximately 4.6% improvement on the minicoco2021 dataset and 4.2% improvement on the remotely sensed dataset VisDrone.

Abstract Speckle reduction is still an intractable task in ultrasound imaging field. Ultrasound speckle is usually described as multiplicative noise with its statistics following a Rayleigh or Gaussian distribution. To employ these two distributions effectively, the authors attempt to describe ultrasound speckle using a Weibull distribution, because it can include the Rayleigh distribution as a special case and also approximate a Gaussian distribution by varying its shape and scale parameters. The authors’ contribution in this paper is to propose a Weibull‐distribution‐based hybrid total variation (WHTV) method to reduce ultrasound speckle. The WHTV energy functional is convex and consists of a new data fidelity term and a new regularization term. The former is derived from the multiplicative Weibull model of ultrasound speckle based on the maximum likelihood criterion. The latter is a new edge‐weighted combination of the first‐ and second‐order total variation, with the advantage of preserving edges while alleviating the staircase effects. The minimization of the WHTV energy functional is implemented by the split Bregman algorithm. Experimental results on synthetic and real ultrasound images have demonstrated not only that the Weibull distribution is a better fitting model for the statistics of ultrasound speckle than other distributions such as Rayleigh, Gaussian, Gamma, and Nakagami, but also that the proposed WHTV method can achieve better despeckling performance than several state‐of‐the‐art variational methods.

Single image super-resolution (SISR) has witnessed great progress as convolutional neural network (CNN) gets deeper and wider. However, enormous parameters hinder its application to real world problems. In this letter, We propose a lightweight feature fusion network (LFFN) that can fully explore multi-scale contextual information and greatly reduce network parameters while maximizing SISR results. LFFN is built on spindle blocks and a softmax feature fusion module (SFFM). Specifically, a spindle block is composed of a dimension extension unit, a feature exploration unit. and a feature refinement unit. The dimension extension layer expands low dimension to high dimension and implicitly learns the feature maps which are suitable for the next unit. The feature exploration unit performs linear and nonlinear feature exploration aimed at different feature maps. The feature refinement layer is used to fuse and refine features. SFFM fuses the features from different modules in a self-adaptive learning manner with softmax function, making full use of hierarchical information with a small amount of parameter cost. Both qualitative and quantitative experiments on benchmark datasets show that LFFN achieves favorable performance against state-of-the-art methods with similar parameters.

Discriminative tracking methods can achieve state‐of‐the‐art performance by considering tracking as a classification problem tackled with both object and background information. As a high efficient discriminative tracker, compressive tracking (CT) has attracted much attention recently. However, it may easily fail when the object suffers from long‐term occlusions, and severe appearance and illumination changes. To address these issues, the authors develop a robust tracking framework based on CT by considering balanced feature representation as well as dual‐mode classifier construction. First, the original measurement matrix of CT works as a dominated texture feature extractor. To obtain a balanced feature representation, they propose to induce a complementary measurement matrix by considering both texture and colour features. Then, they develop two classifiers (dual mode) by using previous and current sample sets, respectively, and subsequently combine them into one ensemble classifier to track the target, which can help to avoid tracking failure suffering from severe appearance changes and long term occlusion. Moreover, they propose a classifier updating schema to prevent the inclusion of unsatisfied positive samples by predicting the occlusions with their ensemble classifier. The extensive experiments demonstrate the superior performance of their tracking framework under various situations.

Biomedical images are usually corrupted by strong noise and intensity inhomogeneity simultaneously. Existing region-based active contour models (RACMs) easily fail when segmenting such images. In the frequency domain, we propose a generalized RACM that presents a new way to understand the essence of classical RACMs whose segmentation results are determined by a frequency filter to extract the proposed frequency boundary energy. Then, we introduce the difference of Gaussians as the optimal filter to exclude strong noise and intensity inhomogeneity effectively. We show superior performance of the model by comparing with six state-of-the-art methods on challenge biomedical images and segmenting an optical coherence tomography image sequence.

This paper presents an adaptive anchor-point selection method for single image super-resolution (SR), which is based upon internal example-based SR model via locality constrained anchored neighborhood regression. The anchor points are fixed in anchored SR methods and are not flexible and customized for different input low-resolution (LR) images. To overcome this defect, we adaptively select anchor points via constructing customized training set for different input LR images, which can be realized by an internal example-based SR method. We introduce a locality-constrained anchored neighborhood regression to learn the relationship between LR space and high-resolution (HR) space. Extensive experimental results demonstrate that the performance of proposed method is competitive with several state-of-the-art SR methods.

Intensity inhomogeneity is one of the major obstacles for intensity‐based segmentation in many applications. The recently proposed kernel mapping (KM) method has exhibited excellent performance on segmenting various types of noisy images while it is not effective to handle intensity inhomogeneity. To overcome this drawback, this study presents a localised KM (LKM) method based on the fact that intensity inhomogeneity can be ignored in a local neighbourhood. The authors’ method first reconstructs the KM formulation of image segmentation in a neighbourhood of each pixel, and then such formulations for all pixels can be integrated together to derive the LKM energy functional. Minimisation of the energy functional is implemented by solving an equivalent convex‐relaxed problem whose optimisation can be quickly achieved via the split Bregman method. Experimental results on two‐phase segmentation and multiphase segmentation demonstrate competitive performance of the LKM method in the presence of intensity inhomogeneity and severe noise.

In view of the existing polygonal approximation algorithm of digital curves can't effectively solve the problem of polygonal approximation constrained by the offset direction, this paper proposes an algorithm of polygonal approximation constrained by the offset direction. First, the offset polygon of the original digital curve is calculated under the control of offset direction and distance. Second, the summation of the squared Euclidean distances between the vertices on the offset polygon and its corresponding segment in the approximated polygon is selected as the fitness function. Finally, under the control of the offset distance and fitness function, this paper implements a PSO-based polygonal approximation algorithm to approximate the offset polygon. Experiments show that the proposed method can not only satisfy the polygonal approximation with directional requirements, but also can greatly improve the operating efficiency.

In this paper, a real-time visual tracking system that delivers superior performance under difficult situations is proposed. The system is based on Histogram of Oriented Gradient (HOG) within the on-line boosting framework. For environmental adaptation, the HOG feature is calculated with blocks of random scale, position and aspect ratio which form a feature pool. The on-line boosting can then select the best distinguishable features from this pool for the robust tracking. The randomness of the blocks guarantees the existence of those features. Three experiments are conducted to highlight different characteristics of this new system. The first experiment proves the validity for the system to be able to pick out the best possible HOG features. The second experiment shows its robustness against bad illuminations and small foreground background difference. The third experiment demonstrates its advancement compared with the Haar-based state-of-the-art system. All those are offered without sacrificing the computation load.

Optical coherence tomography (OCT) is becoming an increasingly important imaging technology in the Biomedical field. However, the application of OCT is limited by the ubiquitous noise. In this study, the noise of OCT heart tube image is first verified as being multiplicative based on the local statistics (i.e. the linear relationship between the mean and the standard deviation of certain flat area). The variance of the noise is evaluated in log-domain. Based on these, a joint probability density function is constructed to take the inter-direction dependency in the contourlet domain from the logarithmic transformed image into account. Then, a bivariate shrinkage function is derived to denoise the image by the maximum a posteriori estimation. Systemic comparative experiments are made to synthesis images, OCT heart tube images and other OCT tissue images by subjective assessment and objective metrics. The experiment results are analysed based on the denoising results and the predominance degree of the proposed algorithm with respect to the wavelet-based algorithm. The results show that the proposed algorithm improves the signal-to-noise ratio, whereas preserving the edges and has more advantages on the images containing multi-direction information like OCT heart tube image.

An audio watermarking algorithm based on psychoacoustic model is proposed. The algorithm can resist the statistics attack. The frequency masking threshold of the audio is calculated through psychoacoustic model. The frequency masking threshold is used to estimate the strength of the watermark signal and determine the location of embedding. The watermark is embedded in a way of resistance statistics attack. The audio watermarking algorithm presented here is blind for no original audio is required when detection. The simulation results show that the watermarked audio is perceptually similar to the original one.

The conventional SVD-based watermarking has been proved to be flawed by protocol attacks. An image watermarking algorithm is proposed based on media hash. Instead of using randomly Gaussian sequence as watermark, a meaningful text message modulated by media hash sequence is used. Theoretical analysis results show that the proposed algorithm solves the problem of protocol attack, while keeping all the advantages of previous SVD-based schemes. Experimental results show that the proposed algorithm is robust and secure. Comparisons with previous algorithms indicate that the performance of the proposed algorithm has been significantly improved.