—Human beings rely primarily on vision to perceive and interact with the external world, with approximately 80% of sensory information input through the visual system. This visual dominance makes the question of "where an individual is looking" not only a key to understanding attention distribution and information processing mechanisms but also a critical factor in optimizing decision-making efficiency and learning outcomes. However, traditional methods for analyzing gaze-related behaviors — such as manual behavioral observation and self-reported evaluation— suffer from inherent limitations: be havioral observation relies on subjective judgment of observers, often missing subtle gaze shifts and failing to achieve real-time tracking; self-evaluation is prone to memory biases and social desirability effects, leading to deviations between reported and actual gaze patterns. These drawbacks highlight the need for a more objective and precise alternative.Gaze estimation, which infers an individual’s visual attention and behavioral intentions by recording and analyzing the spatial position, movement trajectory, and dynamic changes of the eyeball, emerges as an ideal solution. This technology is broadly categorized into model-based (relying on geometric eye models) and appearance-based (using facial/ocular image features) approaches, with appearance-based methods gaining traction due to their non-intrusiveness. Nevertheless, current appearance-based gaze estimation still faces two major challenges: (1) individual differences, such as variations in eye shape, pupil size, eyelid structure, and the presence of glasses, which disrupt consistent feature extraction; (2) environmental interference, including variable lighting, partial facial occlusion, and dynamic head poses, which reduce estimation accuracy. To address these issues, this paper proposes RTACM-Net, a novel gaze estimation network architecture that integrates the strengths of Vision Transformer (ViT) with a multi-scale feature fusion mechanism. Specifically, RTACM-Net employs a lightweight convolutional module to extract local fine-grained features of the ocular region, while leveraging ViT’s multi-head attention mechanism to capture global contextual relationships. This dual-branch design enables the network to balance local feature precision and global context awareness, thereby mitigating the impact of individual differences and environmental noise.Extensive experiments were conducted on two benchmark datasets: MPIIFaceGaze (a large-scale dataset focusing on indoor controlled environments with 21 subjects) and Gaze360 (a challenging dataset covering diverse outdoor/indoor scenes, variable lighting, and large head-pose variations with over 100 subjects). The results show that RTACM-Net : on MPIIFaceGaze, it achieves an average angular error (MAE) of 3.72° ; on Gaze360, it achieves an MAE of 10.46°, Gaze360-Net (11.40°) by 0.94°. These results demonstrate the robustness of RTACM-Net in handling variable individual characteristics and complex environmental conditions. Its practical potential extends to multiple fields: in augmented reality (AR), it can enable adaptive interface rendering; in autonomous driving, it supports dual-task monitorin; in human-robot interaction, it facilitates intuitive service triggering.
—This paper presents an improved version of the DeepLabV3+ network to address issues such as large parameter count, difficulties in mobile deployment, limited receptive field, and insufficient utilization of low-level semantic information in existing deep learning semantic segmentation networks. The main enhancement approach is as follows: we utilize the lightweight MobileNetV2 as the backbone feature extraction network, while an improved multi-scale atrous convolution module (AS-ASPP) and convolutional block attention mechanism (CBAM) are introduced. Tests conducted on the PASCAL VOC 2012 dataset demonstrate that the optimized model retains merely around one-tenth the parameters of the original network, while attaining superior segmentation precision and computational effectiveness. Specifically, it reaches a mIoU of 73.21% and a Precision of 80.56%, with the training time reduced by approximately 50% and the inference speed significantly improved.
—The rapid advancement of mobile technologies has led to increasingly powerful and feature-rich devices, yet this progress has also intensified the challenge of managing energy consumption effectively. Power optimization has therefore become a critical focus in mobile operating systems (OS), aiming to balance performance, functionality, and energy efficiency. As mobile devices integrate more complex hardware components and resource-intensive applications, ensuring sustainable power usage has become essential for improving battery life, user experience, and environmental sustainability. This research explores the fundamental question: How can mobile operating systems intelligently manage hardware and software resources to minimize power consumption without compromising performance or usability? To address this, the study examines key power optimization strategies and mechanisms integrated within modern mobile OS architectures, including Dynamic Voltage and Frequency Scaling (DVFS), power-aware CPU scheduling, Doze and App Standby modes, adaptive display and sensor management, and network optimization. The research also investigates the role of advanced techniques such as context-aware power management and machine learning-based predictive models in achieving dynamic, intelligent energy control. Using tools like Trepn Profiler, PowerTutor, and Android Battery Historian, the study evaluates how power consumption patterns can be analyzed and optimized in real time. The findings reveal that combining hardware-level techniques (like voltage scaling and clock gating) with software-level optimizations (such as adaptive scheduling and contextual awareness) results in significant energy savings while maintaining user satisfaction. Furthermore, the study highlights emerging challenges, including the trade-offs between performance and energy efficiency, the integration of AI for predictive optimization, and the need for sustainability across the device lifecycle. Ultimately, this research demonstrates that power optimization in mobile operating systems is not merely a technical requirement but a cornerstone of sustainable computing. Through intelligent power management, future mobile OSs can achieve greater efficiency, extended device longevity, and reduced environmental impact aligning technological innovation with eco-efficiency and user-centric design.
—To address the challenge of accurate gaze estimation in unconstrained environments susceptible to various interfering factors, this paper proposes AG-HybridNet, an end-to-end gaze estimation model integrating a dual-branch architecture combining CNN and Transformer components. The model employs Swin Transformer as the backbone for global feature extraction while incorporating an enhanced CNN branch dedicated to local feature capture. We introduce the TDConv-Block, which replaces standard convolution with partial convolution integrated with reparameterization technique, significantly reducing computational load and memory access while forming a T-shaped receptive field focused on central facial regions. Additionally, we design Efficient Additive Attention (ED-Attention) that effectively resolves the computational bottleneck in long-sequence processing for Transformers by reconstructing the computational workflow. Comprehensive experiments on MPIIFaceGaze and Gaze360 datasets validate the model's effectiveness. Experimental results demonstrate that AG-HybridNet achieves mean angular errors of 3.72° and 10.82° on MPIIFaceGaze and Gaze360 datasets respectively. Comparative studies with other mainstream 3D gaze estimation methods confirm that our network model can accurately estimate 3D gaze directions while reducing computational complexity.
—Small object detection remains a formidable challenge in computer vision, primarily because conventional models like SSD suffer from two critical limitations: weak semantic information in shallow feature maps and a mismatch between the receptive field and the actual size of small targets. To address these deficiencies, this paper introduces Lite-RFB SSD, an innovative architecture that strategically integrates a lightweight Receptive Field Block (RFB) module into the SSD framework. This module is meticulously reconstructed using depthwise separable convolutions and channel pruning techniques, resulting in a remarkable 62% reduction in parameters. By embedding this optimized module into the shallow conv4_3 layer, the model preserves high-resolution features crucial for small object detection while significantly enhancing computational efficiency. Experimental validation on the PASCAL VOC dataset demonstrates that Lite-RFB SSD achieves an average precision for small objects (APs) of 22.9%, a substantial 4.2% improvement over the original SSD. Furthermore, it operates at an impressive 28 FPS on edge devices, establishing a superior balance between accuracy and efficiency that outperforms competing methods such as standard RFB and MobileNet-SSD.
—Object detection and recognition in drone aerial images hold broad application value, but also present challenges such as large variations in object scales, difficulties in detecting small objects, and occlusions in dense scenes. To address these issues, this paper proposes an improved object detection algorithm based on RT-DETR. First, a Spatial-Channel Collaborative Attention (SCSA) module is introduced into the PResNet backbone network to enhance feature representation and improve detection accuracy. Second, the Content-Aware ReAssembly of Features (CARAFE) upsampling method is adopted in the Hybrid Encoder, which preserves more detailed information of small objects while reducing model complexity, further boosting detection performance. Finally, a modified MFRC3 module incorporating Biphasic Feature Aggregation Module (BFAM) and boundary attention mechanism is proposed to replace the original CSPRepLayer. This enhances multi-scale feature fusion and improves the retention of fine-grained and textural features.Experimental results on the VisDrone2019 datasets show that the improved algorithm achieves an mAP@0.5 of 51.1%, which is 3.1% higher than the baseline RT-DETR model.
—Polar codes, as capacity-achieving error-correcting codes, have become a cornerstone of modern communication systems due to their excellent theoretical performance. Compared with the Successive Cancellation (SC) decoding algorithm, the Belief Propagation (BP) decoding algorithm for polar codes offers advantages such as parallel output and ease of hardware implementation. However, the bit-flipping decoding schemes based on BP still exhibit a significant performance gap in frame error rate (FER) compared to the Successive Cancellation List (SCL) decoding. To address the demand for high reliability and low power consumption in practical applications, this paper proposes an optimized bit-flipping scheme in which the flipping set is constructed using an adaptive genetic algorithm. The proposed method first reduces the computational complexity of the initial BP decoding process by adopting the Offset Min-Sum (OMS) approximation. During the construction of the flipping set, an adaptive mechanism dynamically adjusts the crossover and variational probabilities based on the fitness of individuals in the population. The indices of the information bits are used as individuals in the genetic algorithm, enabling the fitness values to gradually evolve from local optima toward a global optimum. This approach allows for more accurate identification of bit positions prone to decoding errors. For a polar code with a length of 1024 and a code rate of 0.5, the proposed AGA-OMS-BPF decoder achieves approximately 1.3 dB BER performance gain at a BER of 10 −5 compared with the conventional BPF decoder. Simulation results demonstrate that the proposed method effectively reduces the number of unsuccessful BP decoding attempts by constructing a more efficient flipping set, thereby achieving performance gains with reduced computational complexity.
—Tool wear detection in mechanical machining is a critical link for ensuring product quality and improving production efficiency. However, this field faces challenges such as scarce annotated data and interference from complex working conditions, making it difficult to deploy advanced detection models. To address the fundamental mismatch between model capacity and data availability, this paper proposes a novel data-efficient hybrid detection architecture named MD-YOLOV12. This architecture ingeniously integrates the rich general visual representations learned by the self-supervised vision transformer model DINOv3 with the YOLOv12 object detection framework. Specifically, we perform feature enhancement at two key locations: input preprocessing and the middle layer of the backbone network, thereby enhancing the model's perception and recognition capability for tool wear features without relying on massive annotated data. To validate the method's effectiveness, we constructed a specialized tool wear detection dataset containing 8083 high-resolution images, meticulously annotated into three categories: "No Wear," "Moderate Wear," and "Severe Wear." Extensive experimental results demonstrate that the proposed MD-YOLOV12 method surpasses existing state-of-the-art techniques in the tool wear detection task, providing a viable technical pathway for data-efficient industrial vision applications.
—Addressing the common challenges in night vision imagery — poor lighting conditions, low pixel resolution, and diminished contrast — which hinder effective pedestrian feature extraction and result in suboptimal accuracy and real-time performance for night-time pedestrian detection, This paper proposes a deep learning-based night vision pedestrian detection system. Building upon the YOLOv8 object detection algorithm, the model is enhanced by incorporating the CBAM attention mechanism into its network architecture and upgrading the optimiser from SGD to Lion. The system design and development are further tailored to address the specific characteristics of night-time imagery. After experimental simulation verification, the performance of the improved algorithm model has been significantly improved: the overall accuracy is improved by about 2.0%, mAP@0.5 is improved by 1.6%, the average accuracy of IoU threshold 0.5 to 0.95 is improved by about 0.04%, and the F1 Score is improved by 0.64%. The improvement plan proposed in this paper effectively enhances the model's comprehensive identification ability of night vision pedestrians, improves the overall performance of the system, and verifies the correctness and validity of the research.
—Driven by deep learning advances, speech processing systems such as automatic speech recognition (ASR), source segregation, noise suppression have achieved significant performance improvements. However, traditional training strategies, particularly static mini-batch selection, often overlook the dynamic variations in data complexity and model convergence behavior, resulting in ineffective training efficiency and limited model accuracy. To tackle this limitation, we introduce a novel training paradigm called Dynamic Micro-block Optimization (DMBO). The method introduces a fine-grained sampling mechanism by partitioning the training set into smaller units called “micro-blocks,” which are dynamically updated during training based on real-time characteristics such as sample loss, gradient diversity, and utterance complexity. Four sampling strategies—loss-weighted, gradient-diversity, gender-based, and accent-based—are designed to self-adjust the composition of training data. The DMBO framework is implemented using Connectionist Temporal Classification (CTC) and Long Short-term Memory (LSTM) networks for end-to-end speech recognition. Experimental evaluations on the VCTK datasets demonstrate that the proposed method significantly accelerates convergence and improves model accuracy. Specifically, the gender-homogeneous strategy reduces the Label Error Rate (LER) by 9.0% compared to standard mini-batch training, while the accent-heterogeneous strategy achieves a 9.2% absolute LER reduction. These results confirm that dynamic optimization at the micro-block level enhances the efficacy of deep learning models in speech processing tasks, and the experimental outcomes are consistent with theoretical expectations, validating the effectiveness and correctness of the proposed approach.
Aiming at the significant deficiencies of traditional traffic signal control algorithms in multi-intersection collaboration, unexpected event response and system generalization ability, this paper proposes an intelligent traffic signal control method that integrates a bidirectional gated recurrent unit BGRU with deep reinforcement learning DRL. The method adopts BGRU to model the historical traffic flow data in time sequence and accurately predict the traffic state; and based on this, it constructs deep Q-network intelligences to dynamically optimize the multi-intersection signal timing strategy. The experimental validation on SUMO simulation platform shows that the proposed method effectively improves the control performance. Compared with the traditional fixed-cycle control and adaptive control methods, the proposed method reduces the average vehicle waiting time by 58.8% and 42.2%, and improves the intersection access efficiency by 83.8% and 52.4%, respectively. The study provides new ideas for building an efficient and intelligent urban traffic management system.
Uncertainty in decision-making processes presents a critical challenge for autonomous agents, often leading to suboptimal or erroneous policies. This paper addresses two prevalent yet distinct types of uncertainty that significantly degrade agent performance: fuzzy uncertainty, stemming from ambiguous task boundaries, and gray uncertainty, arising from noisy or incomplete state observations. To tackle these challenges, we propose the Dual-Task-State Inference (DTS-Infer) method, a novel framework that leverages variational inference within an off-policy reinforcement learning structure. DTS-Infer utilizes a dual-network architecture to explicitly disentangle and resolve these uncertainties: (1) a task inference network learns a latent distribution over tasks from historical data to disambiguate task goals, thereby solving the fuzzy uncertainty problem ; and (2) a state inference network captures robust latent features of the current state to overcome corrupted sensory input, thus addressing gray uncertainty. Extensive experiments on continuous control benchmarks demonstrate that DTS-Infer significantly outperforms state-of-the-art algorithms. For instance, in the Half-Cheetah-Fwd-Back environment, DTS-Infer achieved a final average reward of 1612.61, representing an 18.9% improvement over the PEARL algorithm. Furthermore, ablation studies confirmed that our inference modules contribute to an 80% increase in average reward over a standard TD3 baseline, highlighting the method's effectiveness in enhancing the robustness and adaptability of intelligent agents.
In UAV-captured images, the high density of objects and the large proportion of small targets pose significant challenges to YOLO-based object detection algorithms. This study presents an enhanced object detection framework derived from the YOLOv10s architecture, aiming to achieve superior detection accuracy. First, an Adaptive Progressive Feature Unification (APFU) module is proposed to effectively integrate multi-level feature representations, ensuring a balanced fusion of high-level semantic information from low-resolution features and fine-grained spatial details from high-resolution features. Second, a Feature Enhancement and Attention (FEA) module is introduced to adaptively recalibrate feature responses, emphasizing informative features while suppressing irrelevant noise and interference. Finally, based on these modules, the APFU-YOLOv10 network is built to effectively improve the network's perception ability of objects at different scales. Experimental results on the VisDrone dataset demonstrate the superior performance of the proposed algorithm: mAP@0.5 increased from 42.6% to 43.5%, a relative improvement of approximately 2.11%; mAP@0.5:0.95 improved from 25.4% to 26.2%, a relative increase of about 3.15%; recall improved from 0.410 to 0.416, further reducing missed detections and enhancing object coverage. The method achieves significant improvements in detection accuracy under medium to high IoU thresholds, validating the effectiveness of multi-scale feature fusion and adaptive attention mechanisms in small object detection for UAV imagery.
Satellite communication has served as the foundation for television, radio, and telephone transmission for more than a century. These communications function at extremely high frequencies, primarily 6 GHz for uplink and 4 GHz for downlink. Satellite antennas installed on residences convert these high-frequency signals downward to make more efficient use of them. Frequency down-converters are commercially known as Low-Noise Blocks (LNBs). LNBs are responsible for receiving, amplifying, and then down-converting these microwave signals to a lower range of intermediate frequencies. This down-conversion is essential as it enables the signal to be transmitted through relatively inexpensive coaxial cables, in contrast to the costly and impractical waveguides that would be necessary for transmitting the original microwave signals. This paper addresses the design of the three primary components that constitute a frequency downconverter: the Low Noise Amplifier (LNA), the Local Oscillator (LO), and the Frequency Mixer. The intermediate frequencies required for satellite applications typically range from 75 MHz to 900 MHz. This study designs a frequency down-converter that generates an intermediate frequency of 100 MHz. For an input radio frequency of 1 GHz, the oscillator will be designed to operate at a center frequency of 0.9 GHz.
To address the challenges of low reconstruction accuracy and insufficient model generalization in image super-resolution (ISR) under complex degradation scenarios, this paper proposes an improved method that integrates generative adversarial networks (GAN) and vision transformers (ViT). First, in the generator module of Real-ESRGAN, some residualin-residual dense blocks (RRDB) are replaced with ViT modules, leveraging the self-attention mechanism to enhance global feature modeling. This enables the model to better capture global information while preserving local details in complex scenes. Experimental results demonstrate that the improved model achieves PSNR gains of 0.59dB/0.45dB and SSIM improvements of 0.018/0.056 in ×2/×4 upscaling tasks on the Urban100 dataset, while also exhibiting excellent performance on benchmark datasets such as Set14. This method significantly enhances image reconstruction quality under complex degradation conditions, providing an effective technical solution for practical applications such as security surveillance, remote sensing monitoring, and target reconnaissance.
As the technologies of virtual reality and augmented reality rapidly advance, the demand for high-quality 3D models has been growing exponentially. However, the Multi-View Stereo Network (MVSNet) for 3D reconstruction has faced issues with the inaccurate extraction of global image information and depth cues. In response to these challenges, this paper presents enhancements to MVSNet. First, the self-attention mechanism is introduced to enhance MVSNet's ability to capture global information in images. Second, a residual structure is added to mitigate the accuracy loss caused by the downsampling and upsampling of feature maps during the regularization process of cost volume, thus ensuring the integrity of information and transmission efficiency. Experimental results indicate that, in comparison with the original MVSNet, the SelfRes-MVSNet reduces the error rate by 1.3% in terms of overall accuracy and completeness, thereby improving the reconstruction effect from 2D images to 3D models.
Image inpainting represents a sophisticated methodology within the domain of computer vision, whose core objective is to programmatically restore occluded regions or eliminate undesired elements from digital imagery. This process endeavors to reconstruct visual continuity such that the resulting image exhibits both perceptual naturalness and structural completeness. Image inpainting has gradually become a hot field in computer vision. It is used in film processing, watermark removal, photo processing, and other fields. Traditional image inpainting methods use adjacent pixels of the missing area for filling, which not only incur high computational costs but also suffer from ghost artifacts and blur. With the emergence of large-scale datasets, deep learning-based image inpainting methods have been successively proposed, significantly improving restoration quality. However, the current state-of-the-art methodologies continue to demonstrate suboptimal performance when confronted with images featuring extensive occluded domains. Additionally, technological advancements in related image fields bring new opportunities and challenges to image inpainting. This paper discusses three aspects: (1) a review of relevant datasets for image inpainting, (2) a detailed description and summary of state-of-the-art methods, and (3) an introduction of evaluation metrics with performance comparisons of representative approaches. Finally, we address existing challenges and future opportunities in this field.
This research investigates the development of a custom hybrid operating system (OS) for a Mars rover experimental prototype using the Raspberry Pi platform. Focusing on operating system optimization, the work enhances computational efficiency, real-time responsiveness, and AI integration. Key innovations include overclocking (boosting CPU performance by 28%), custom threading (reducing task scheduling latency by 22%), and networking improvements for stable remote operation. Codec refinements and framework adaptations improved real-time video analysis throughput by 30%. Integration of a Power-over-Ethernet (PoE) HAT enhanced thermal regulation and stabilized system runtime. Experimental results show the customized OS effectively supports intensive tasks such as image processing, sensor data acquisition, and edge AI workloads. The findings demonstrate a scalable, modular OS framework for real-time vision systems and intelligent robotics in resource-constrained environments.
This study investigates the optimization of broadband communication channel capacity through an integrative information-theoretic framework. Leveraging Shannon’s theory, it examines fundamental constraints such as bandwidth limitations, channel noise, modulation techniques, error correction mechanisms, and adaptive systems. A comprehensive literature review of 118 articles identified 18 critical enablers, which were evaluated by domain experts. The Fuzzy DEMATEL method was employed to prioritize enablers based on interdependencies and influence. Results indicate that Security Considerations, Channel Access Protocols, and Propagation Characteristics exert the most significant impact on capacity optimization. The findings offer a structured decision-making model for stakeholders, enabling efficient allocation of technological, infrastructural, and human resources. By bridging theoretical principles with practical implementation, this research provides actionable insights for academic researchers and industry practitioners in designing robust, high-capacity broadband systems. The integrative modeling approach advances the application of information theory in modern communication networks, supporting informed technology adoption and system integration.
It’s highly crucial to divide up medical photos correctly in order to make diagnoses and plan treatments. Convolutional Neural Networks (CNNs) are very good at picking up local information, but they have problems with long-range dependencies. On the other side, Vision Transformers (ViTs) are good at modeling global context, but they need a lot of computer power and labeled data. To get surrounding these difficulties, we establish PSwinUNet, a hybrid CNN-Transformer system based on a partially supervised learning the structure. Adding a SwinTransformer block to a U-shaped structure makes PSwinUNet better at learning internationally semantics and up-sampling. It also uses a polarized self-attention mechanism in skip connections to keep spatial information from getting lost when the image is downsampled. PSwinUNet does a better job than the best gets closer that are currently accessible when tested on the BUSI, DRIVE, and CVC-ClinicDB datasets. For instance, it earned Dice Similarity Coefficient (DSC) scores of 0.781, 0.896, and 0.960 based on the BUSI data set with 1/8, 1/2, and entire labeled information, respectively. These scores are substantially better than those of the old UNet and UNet++ models.
