In recent years, it has become increasingly important to utilize the entire network links more effectively to avoid traffic congestion for Internet Service Providers (ISPs), where link installation costs are high. As a promising approach to address this issue, multipath routing, which distributes traffic across multiple reachable paths to the destination, has getting attention. In multipath routing, congestion can be avoided by using other paths and balancing path loads even if a path is congested. Conventionally, realizing load-aware multipath routing has required both the collection of load metrics to track dynamically changing path loads and the distribution of traffic at an appropriate ratio with fine-grained traffic units such as flowlets. However, in ISP networks, existing methods may fail to balance path loads due to the large path delay and the variation in flow bit rates. In this paper, we propose a novel traffic balancing method suitable for ISP networks. In the proposed method, we first derive a target bandwidth for each path to equalize congestion levels of all paths in multipath, and then decide the distribution ratio by feedback control. In addition to this, the proposed method adopts modified flow-level traffic distribution, which makes flows reselect their paths at a certain time intervals. These approaches enable to balance traffic more evenly in ISP networks than conventional methods. Through network simulations using network topologies assuming ISP networks, including SINET6, we demonstrated that the proposed method can reduce the average flow completion time (FCT) by 16.0%, 44.5%, and 58.4% compared to ECMP, which performs naive traffic distribution, CONGA and W-ECMP, which achieve advanced traffic distribution.

State Machine Replication (SMR) is a fundamental technique for building fault-tolerant distributed systems with strong consistency. Rabia is an SMR protocol that simplifies implementation design through a randomized consensus algorithm. Our analysis reveals a design limitation of the Rabia protocol: under partial network partitioning, replicas can develop inconsistent queue states, leading to a livelock state. We present Qsync, which enhances Rabia’s fault tolerance through queue state synchronization mechanisms while preserving its implementation simplicity. Experimental evaluation shows that Qsync maintains stable performance under partial network partitions where the original Rabia throughput drops to zero.

Driven by the strong demand for enhanced performance in High-Performance Computing (HPC), Coarse-Grained Reconfigurable Architectures (CGRAs) are promising technologies that offer high performance even under power consumption constraints. Performance on CGRAs is significantly influenced by loop unrolling, a technique that increases computational parallelism by utilizing more processing elements in CGRAs. Determining the optimal loop unrolling factor is challenging in applications with multiple loops. This paper presents a case study demonstrating the determination of optimal loop unrolling factors for an application based on the Lattice Boltzmann Method (LBM). Because the application’s process exceeds the capacity of a single CGRA, this paper proposes a method for partitioning the process to fit the CGRA’s resources using integer linear programming (ILP). Finally, this paper provides a performance estimation of the CGRAs runtime and demonstrates the effectiveness of CGRAs for HPC.

The coarse-grained reconfigurable architecture (CGRA) has been attracting significant attention as an energy-efficient accelerator. Recently, many applications require significant computational power, and CGRAs are expected to meet this demand. In such fields, CGRAs are utilized to execute computationally intensive programs in the innermost loop body, often called a “kernel”. They generally consist of a two-dimensional array of processing elements (PEs) interconnected in a configurable manner, and the data transfer between PEs is configured accordingly. Running a kernel on a CGRA requires a mapping process that generates CGRA configurations to match the kernel program. This mapping is time-consuming and can hinder developer productivity. We therefore propose a fast mapping method that leverages the architecture’s characteristics, namely, its routing capabilities, to reduce mapping time. We define a heuristic cost function for routing that guides the mapper toward better mapping results. We demonstrate that our mapper is fast enough for practical software development and can provide sufficiently robust mapping results.

This paper proposes a Field-Programmable Gate Array (FPGA) accelerator for Vision Transformers (ViTs) with quantization and look-up-table (LUT) based operations. First, two improved quantization methods are proposed, achieving comparable performance at lower bit-widths. Furthermore, linear and nonlinear units’ designs are proposed to support diverse operations in ViTs models. Finally, the LUT-based accelerator design is implemented and evaluated. Experimental results on the ImageNet dataset demonstrate that our proposed quantization method achieves an accuracy of 80.74% at 2-bit width, outperforming state-of-the-art Vision Transformer quantization methods by 0.1% to 0.5%. The performance of the proposed FPGA accelerator demonstrates a higher energy efficiency, achieving a peak energy efficiency of 7.06 FPS/W and 246 GOPS/W.

In the membrane computing, most of the proposed algorithms for computationally hard problems use an exponential number of membranes, and reduction in the number of membranes must be considered in order to make the membrane computing a more realistic model. In the present paper, we propose an asynchronous P system using improved branch and bound to solve the minimum Steiner tree. The experimental results show the validity and efficiency of the proposed P system.

Interconnection networks are inevitable in parallel computers. Effectiveness in parallel execution is largely affected by the interconnection network as a communication performance. Especially, collective communication is important since it is frequently executed in parallel programs. To improve the performance of collective communication, one of the promising methods is packet scheduling. This paper addresses a lazy method for packet scheduling. The proposed method is based on an evolutionary idea to find hopeful candidates for injection delays and improvement methods. Preliminary evaluation results reveal that the proposed method outperforms the existing method.

Automatically reconstructing structured 3D model of real-world indoor scenes has been an essential and challenging task in indoor navigation, evacuation planning and wireless signal simulation, etc. Despite the increasing demand of updated indoor models, indoor reconstruction from monocular videos is still in an early stage in comparison with the reconstruction of outdoor scenes. Specific challenges are related to the complex building layouts which need long-term video recording, and the high presence of elements such as pieces of furniture causing clutter and occlusions. To accurately reconstruct the large-scale indoor scenes with multiple rooms, this paper designs a large-scale indoor multiple room 3D reconstruction pipeline to explore the topology relation between different rooms from long-term monocular videos. Firstly, semantic door detection based video segmentation is proposed to segment different rooms in video for individual reconstruction to avoid global mismatching noise, and 3D temporal trajectory is proposed to connect different rooms in spatial domain. Secondly, 3D Hough transform and Principal components analysis are utilized to refine the room boundary from reconstructed point clouds, which contributes to the accuracy improvement. Further, an original long-term video dataset for large-scale indoor multiple rooms reconstruction is constructed, which contains 12 real-world videos and 4 virtual videos with 30 rooms. Extensive experiments demonstrate that the proposed method reaches the highest performance of the 3D IoU at 0.70, room distance accuracy at 0.87, and connectivity accuracy at 0.67, which is around 39% better on average compared with various state-of-the-art models.

This study proposes a knowledge-based handcrafted building region extraction algorithm that can accurately identify the building and its background from street image at pixel level. The proposed algorithm leverages a customized patch-based graph cut inspired by human visual perception mechanisms. At the patch-based graph cut, the similarity of patches is measured by the cutting-edge deep neural networks (DNNs). The graph settings are based on the knowledge that buildings are captured at the center of the image owing to their main subject. Our experiment, which employed 300 images included in well-known open dataset, demonstrated that the proposed method employing GrabCut for a pixel-level segmentation significantly increased the comprehensive accuracy of building region extraction, which is measured by intersection over union (IoU), by 12.29% or more compared with the conventional knowledge-based method using color segmentation. This stems from the fact that the proposed method presents the more accurate building and background candidates by 8.57% or more. In addition, the GrabCut-based proposed method represented a similar accuracy to the state-of-the-art DNN-based semantic segmentation based on a transformer architecture. Further comparisons and discussions are provided in this paper to clarify the effectiveness of the proposed method.

In recent years, zero-shot learning-based haze removal methods using a single image have been proposed and have gained attention for their effectiveness. However, methods that fuse near-infrared (NIR) and color images have not been sufficiently studied. This paper presents a haze removal method based on zero-shot learning that fuses NIR and color images. The proposed method consists of two steps: haze removal and edge fusion. In the first step, the atmospheric scattering model is adapted to remove haze from NIR and color images. This step restores colors in the color image and enhances edges in the NIR image. In the second step, a new method is introduced to fuse haze-removed NIR and color images. This method preserves the natural color and the luminance of the color image and effectively uses the edges of the NIR image. Specifically, a weight map is generated to adjust for luminance changes and is added to the NIR image. The adjusted NIR image is then multiplied by the lightness image to restore the edges. This process allows for a natural fusion of NIR and lightness images and an effective fusion of detailed edges. Our qualitative and quantitative evaluations demonstrated that our method can restore color and edges more naturally than the conventional methods. Furthermore, it was shown to be effective even for strong haze images.

We propose a non-contact heart rate (HR) estimation method that models weak physiological blood volume pulse (BVP) signals and strong noise signals caused by background illumination. Our method integrates BVP signal extraction based on a physiological model and a flexible RGB/NIR integration scheme based on an illumination model in a unified manner. This unified framework enables accurate extraction of the BVP signal while suppressing noise derived from ambient light, and thus improves HR estimation performance. We demonstrate the effectiveness of our method through experiments using several datasets, including various illumination scenes. Our code will be available on https://github.com/kosuke-kurihara/PhysIllumHR.

We propose a JPEG format-compliant encryption method in the quantized discrete cosine transform (QDCT) domain for texture protection, called Prediction Error-Propagated Encryption with Modulo Operator (PEPE-MO = WPE-MO, by pronouncing ‘W’ as ‘double’). In the QDCT domain, both the direct current (DC) coefficients, which contain structure information, and alternating current (AC) coefficients, which contain texture information, are encrypted with newly placed prediction, encryption, and reconstruction modules. The resulting propagated prediction error reinforces texture protection. To ensure JPEG compatibility, WPE-MO incorporates a modulo operator into the prediction and reconstruction modules, circulating coefficients within the JPEG-encodable value range. Additionally, to balance attack resilience and coding efficiency, two adjustable parameters are introduced: random length interval (RLI) and random step size (RSS). Experiments on JPEG image encryption demonstrate that WPE-MO exhibits high attack resilience with minimal degradation in coding efficiency. In particular, WPE-MO resists ciphertext-only attacks, including brute-force and replacement attacks, with approximately 19.55% degradation in coding efficiency, as measured by the Bjøntegaard-delta rate, through careful selection of RLI and RSS.

Deep neural network (DNN) pruning is a popular method for accelerating computations in DNNs by removing unimportant parameters. Among pruning methods, tile-wise pruning (TWP) achieves significant acceleration with minimal pruning loss. However, TWP suffers from load imbalance when important weight elements in the matrices of the DNN are unevenly distributed. To address this issue, we propose adaptive tile pruning (ATP), an integrative solver for building sparse DNNs with controllably balanced workloads. ATP comprises three components: hierarchical tile pruning (HTP), split-tiled sparse matrix multiplication (STSpMM), and adaptive pattern selection (APS). HTP constructs sparse matrices with evenly distributable workloads while preserving DNN model accuracy. STSpMM efficiently handles HTP-generated sparse matrices on GPUs by splitting and redistributing large workloads. APS dynamically selects pruning patterns for HTP and grid sizes for STSpMM based on the problem sizes in the targeted DNN. We evaluated our approach on pruned ResNet-18 and ResNet-34 models using ImageNet, and BERT-Small on the question-answering natural language inference (QNLI) task. Results demonstrate that models accelerated by ATP achieve greater acceleration than previous methods while maintaining accuracy for inference.

For linear least squares estimation of parameters in wireless sensor networks, we focus on construction of the best subset of sensors that minimizes the estimation error which requires computation of the inverse of large matrices. We manipulate the estimation error based on the LU factorization, resulting in the factored triangular matrices, the inverse of which can be iteratively obtained without large matrix inversion. We then derive an analytic selection rule in a greedy manner which facilitates a fast selection process. We also discuss the complexity of different selection methods with an emphasis on a reasonable complexity of the proposed method. We finally validate the merit of the proposed algorithm through numerical experiments in the aspect of estimation performance and complexity as compared with previous methods.

The docker container-based virtualization is becoming mainstream in cloud computing due to its potential benefits, such as a lightweight resource footprint and quick deployment. However, docker containers can suffer from a lack of stability in provisioning per-container storage space, which can lead to undesirable consequences, such as abrupt application termination requiring execution restart or substantial data loss. In this paper, we propose an I/O storage scheme(mSEM) to enhance the storage reliability and stability of docker containers by dynamically enlarging the data reservoir for each container through the effective data path redirection to our extensible storage space. We measure the performance of our method, while comparing its performance to the baseline docker and kubernetes using three benchmarks: filebench, postmark and vdbench. The results show that our method produces three times higher I/O bandwidth than both baselines under storage shortage conditions. More importantly, even when there is no available storage space left and the baseline stops execution, our method can continue application execution with no severe performance degradation.

In recent years, security measures for IoT devices have become more important. Fuzzing of IoT devices is an effective way to find unknown vulnerabilities. In IoT device fuzzing, a large number of test cases are generated based on a set of initial seeds and they are sent to a target device to monitor its behavior, in which a device crash means that a vulnerability has been discovered. However, generating a set of initial seeds is difficult because technical knowledge in security and adaptation to various IoT devices are quite required. In this paper, we propose a method to generate initial seeds for IoT device fuzzing effectively utilizing a large language model (LLM). The proposed method efficiently generates initial seeds for fuzzing the target IoT device by inputting only the type of IoT device, communication logs, and the name of the vulnerability to be inspected into an LLM, with no specific technical knowledge in security. Experimental results of applying the proposed method to two types of IoT devices show that the proposed method detected the first crash in 0.40 seconds and 0.47 seconds from the start of fuzzing, respectively, and after 24 hours of fuzzing, it detected all the crashes due to null pointer exception and buffer overflow that could not be detected by fuzzing with the initial seeds generated manually.

Dynamic meshes reasonably represent time-varying 3D objects, but compression is required due to the large amount of data involved. One efficient framework decomposes a dynamic mesh into a base mesh and displacements using decimation and subdivision. The displacements are converted to levels by wavelet transforms and quantization, and they are coded by arithmetic coding. The levels of the current frame are predicted from the reference frame, and only the residuals are coded. However, the residual tends to be large since the coefficients of each frame are quantized before performing inter prediction. In this paper, we propose a method of quantizing the residuals obtained after applying inter prediction in order to reduce the amount of required data. The experimental results show that the proposed method improves coding efficiency (BD-Rate: -0.3%) and that the reconstructed mesh has no quality degradations.

An adaptive spatial approximation algorithm is proposed while minimizing false contouring based on local image characteristics considering human visual systems to save the amount of image frame memory and power for image data transmission. The proposed algorithm is generalized by using block-based k-means clustering to categorize image blocks with the same characteristics, and the amount of approximation is evaluated for each cluster. Two different image quality standards are maintained to maximize image approximation while maintaining the required image qualities. The proposed algorithm can reduce frame buffer memory up to 35.11% on average.

Precisely detecting obstacles on the track is critical to the safety of railway transportation. However, existing track obstacle detection methods suffer from issues of low accuracy, slow speed, and high complexity, which are not qualified for real-time demand and low-resource constraints. This paper proposes a novel Railway Obstacle Detection (ROD) method named ROD-YOLO, striking a good trade-off between performance and efficiency. Firstly, we design a multi-scale Feature Enhancement Module (FEM), utilizing convolutions with different dilation rates to extract fine-grained features from different layers. Secondly, to improve detection speed, we propose the SPPCSPC-F spatial pyramid pooling module, which reduces the number of convolution units, the size of pooling operations and the dimensions of feature concatenation. Additionally, we incorporate the Large Selective Kernel (LSK) Attention to filter out interfering information and focus on important local features. Comprehensive experiments are conducted on a real-world dataset consisting of 12,270 images, aiming to verify the feasibility of object detection methods in complex railway environments. Results show that ROD-YOLO outperforms state-of-the-art one-stage and two-stage object detection methods, achieving 96.3% in precision, 91.4% in recall, and 96.6% of mAP at 0.5 IoU threshold. Compared to the most light-weight baseline (YOLOv8n), our method improves the mAP50 and inference speed by 7.93% and 72.42%, respectively, with only 36.19% growth in parameter size. Moreover, ROD-YOLO shows strong generalization ability on four cross-domain datasets, including a remote sensing image dataset and a traffic sign dataset. In conclusion, the proposed ROD-YOLO algorithm demonstrates remarkable performance in detecting track obstacles, provides valuable practice for deployment of object detection models in resource-constrained and security-crucial systems.

Tibetan text recognition plays a key role in preserving the Tibetan language, religion, and traditions. While text recognition has made progress for high-resource languages, handwritten Tibetan character recognition remains difficult due to limited data and the lack of public large language models. Most existing datasets focus on printed or historical documents, as well as online handwriting data, but there are still few large offline handwritten Tibetan datasets. To solve this problem, we construct TibHCR, a large-scale offline handwritten character recognition dataset for the Tibetan language. To increase the diversity of the linguistic and font styles, more character categories and participants from 5 provinces in China are included. To collect and label the data efficiently, we introduce a grid sheet design, reducing manual annotation to just 1% of the samples. This design then allows for automatic data processing to extract each character sample and its corresponding label. The resulting TibHCR dataset contains 141,698 samples from 235 Tibetan writers, covering 47 character classes. We evaluate TibHCR using two recognition models: a convolutional recurrent neural network (CRNN) and a cross-lingual fine-tuning method, on a Chinese pretrained model using the PP-OCRv4 architecture to adapt Tibetan data. The results show that both models can recognize handwritten Tibetan characters efficiently, with an accuracy of 99.48% for CRNN and 99.70% for the fine-tuning method. The TibHCR dataset is publicly available at https://huggingface.co/datasets/qixiaoke/TibHCR.

Owing to the inherent sparsity of the user-item interaction matrix, the majority of existing collaborative filtering-based recommendation algorithms predominantly focus on the explicit interactions between users and items, thereby neglecting the complex interdependencies among items and users. This oversight results in a suboptimal representation of user and item characteristics, ultimately leading to a diminished quality of recommendations. To address this limitation, we proposed a novel recommendation algorithm, the Dual Co-occurrence Convolutional Neural Network (DCoCNN). DCoCNN innovatively integrates three pivotal components: user-item interactions, user-user co-occurrences, and item co-occurrences, leveraging the powerful feature extraction capabilities of CNN to train and refine latent features. Since items or users often emerge in pairs, DCoCNN thoroughly explores the intrinsic relationships among items or users, compensates for the lack of item-user interaction behaviors, and enables the trained latent features to contain more effective co-occurrence information, thereby enhancing model performance. The experimental results show that DCoCNN can effectively capture effective information between items or users, effectively mitigate the deficiencies with non-co-occurrence and single co-occurrence models, and improve recommendation quality.

With the development of society, people get news more and more frequently from online media. Under such circumstances, fake news has become a major social problem. Most of the existing fake news detection works focus on the extraction of identification information. However, how to deal with domain shift problem is still a challenge. In this paper, we propose an approach called Joint Domain-specific and Domain-shared Learning (JDDL) for multi-domain fake news detection. It mainly consists of three modules: (1) The multi-domain feature extraction module, which extracts domain-specific features and domain-shared features, respectively; (2) The feature fusion module, which employs Graph Attention Network (GAT) to further extract features, and then fuses the output features; (3) The domain adversarial discrimination module, which designs the domain discrimination loss to confuse classifier and make it be unable to distinguish which domain the news belongs to. Experiments on English dataset show that the JDDL outperforms state-of-the-art methods.