Traditional facial animation models do not particularly stress the naturalness of imitating the dynamic facial expressions, including eye movements, dynamic eyebrow deformation, and head motion naturalness. Hence, in this paper, a speech-driven talking-face synthesizer (SDTS) is proposed for generating the dynamic talking video of a given static face for semantically mimicking the speech of any real person. The SDTS can lead the static digital-twin face to vividly mimic the expressive motions of the face and lip-synced mouth of various speakers with a personalized accent with high distinctiveness. The SDTS framework has two stages. In the first stage, one branch, termed the dynamic fused-features generation module (DFGM), contains a cross-modal speech-facial fusion module (CSFF) and a temporal convolutional network (TCN). The CSFF is the core to seamlessly align the speech features and facial features. The second branch is the self-designed adaptive identity extractor (AIE), where a series of residual blocks using partial batch normalization unit (PBN-ResNet blocks) and the residual blocks with the squeeze-and-excitation unit (SE-ResNet blocks) are cascaded to precisely capture the key features of the face in a static reference image. In the second stage of SDTS, the diffusion model termed diffusion-based rendering model (DIRM) is applied to generate the high-resolution video reconstruction with the consistencies of appearance and emotion via fusing the driving speech features and the referred facial features. The extensive experiments demonstrate that SDTS can significantly promote lip-synchronization, enrich the upper facial expression, and exhibit the naturalness of the head movements. Moreover, the SDTS can steadily maintain the facial identity consistency and the facial expression coherence for varying speaking speeds and emotions. Hence, it can attain less than 5.26 FID, 0.72 LSE-D, and 0.56 LME than the StyleTalk model, which is a well-known talking-face synthesis model.

The development of advanced soft robotics is crucial for delicate medical operations, as traditional rigid robots lack the flexibility required for such procedures. However, the application of nonlinear time-varying dynamics mechanisms to achieve accurate deformation control still faces challenges. In this study, we address the design of a surgical-tool deformation controller for minimally invasive procedures and propose a collaborative framework that integrates a variational autoencoder (VAE) with a two-dimensional time-varying model predictive controller (MPC). In our primary simulation, compared with the proportional integral derivative (PID), MPC results in improvements by reducing the convergence time by 50%, lowering deformation error by 75%, producing smoother control inputs, and consuming one-sixth the energy. Moreover, MPC demonstrates strong adaptability, stability, and high precision in dynamic operating conditions, thereby ensuring surgical safety, particularly when subjected to discrete external forces. To reduce the computational load in real-time optimization, the VAE is trained offline using MPC data, and subsequently generates the optimal control sequences while enhancing interpretability. The proposed VAE-enhanced soft-robot deformation controller, termed VAI-SDC, can achieve approximately 17% greater energy savings and higher control accuracy than standalone MPC. The experimental results demonstrate that VAI-SDC can attain precise, energy-efficient deformation control of surgical instruments in time-varying environments. As the utilization of VAI-SDC minimizes the invasiveness of surgery, it promises reliable robotic assistance for the safety of clinical applications.

Deploying high-fidelity deep learning models for robotic harvesting on edge platforms creates a fundamental conflict: the inference latency often exceeds stringent real-time control deadlines, resulting in the critical "stale data" problem. We present a cohesive edge computing system that resolves this conflict through two core contributions. First, we introduce M2PE, a high-fidelity perception engine that achieves an accuracy of 97.4

In long-horizon multi-party conversations, human-centric AI agents face a persistent structural problem: similarity-based retrieval may fail to reconnect semantically dispersed fragments of the same evolving event. This problem severely weakens causal continuity and multi-hop context recovery. To improve attribution trust and reduce structural erasure, we propose MemLoom, a dual-loop causality-traceable retrieval framework that organizes conversational history as an event memory graph. MemLoom decouples latency-sensitive online interaction from off-peak structural curation through online event formation, sentence-level buffering, asynchronous neuro-symbolic graph synthesis, and bounded dual-stream retrieval. Evaluations across QMSum, LoCoMo, and the synthetic causal diagnostic suite (SCDS) support the structural utility of MemLoom. For LoCoMo, under our unified local evaluation setup, MemLoom shows favorable temporal and multi-hop reasoning results (J = 65.77 and 58.14) relative to contemporary agentic baselines, such as Mem0, Zep, and A-Mem. For SCDS, within a controlled diagnostic setting, it recovers demanded causal chains more reliably than GraphRAG (SCR = 0.72 vs. 0.35) and maintains stronger answer-level auditability (AA = 0.80 vs. 0.50). This is achieved with a bounded online P95 latency of 1.67 s. These results indicate that asynchronous dual-loop stewardship has practical value for causality-traceable, event-centric conversational memory in multi-party settings.

In delivery tasks, outdoor logistic robots are more versatile than those designed for indoor use only. In this study, we developed a low-cost outdoor autonomous mobile robot (LOAMR) with a tracked design, capable of operating effectively in both outdoor and indoor environments. Our LOAMR is meticulously constructed with all the controlling ingredients, requisite sensors, hardware complement, and circuit boards on the self-made robot body. In LOAMR, the harmonic steering controller, named PIC-FOC controller, is self-built by integrating proportional-integral control (PIC) and field-oriented control (FOC). The integrated controller can ensure the movement stability of LOAMR outdoors with the proposed outdoor navigation technology. This technology can perform navigation effectively in various outdoor environments without the time-consuming SLAM creation and mapping mechanism. Finally, the PIC-FOC controller, SiLU-improved Yolov7-tiny object detector, GPS navigator, and ResNet34 walkable-area detector are combined to attain the outdoor march amid the buildings. Meanwhile, the detected RGB image of SiLU-improved Yolov7-tiny is coupled with the depth map of the TOF camera and the GPS-based Google Map contributes to the obstacle avoidance continuously. The self-constructed LOAMR can be considered a highly cost-effective outdoor-conveyance AI apparatus.

LSTR (Lane Shape Prediction with Transformer Network) is a curved road shape detection model based on the Transformer architecture. It provides robust road shape detection capabilities and a sound context consistency structure. This study introduces RE-LSTR (Robustness-Enhanced LSTR), an improved variant designed for greater stability and reliability in autonomous driving. RE-LSTR addresses challenging conditions including low-light environments, severe occlusion, degraded markings, and heavy shadows. It integrates a Shuffle Attention mechanism to enhance inter-feature dependencies and adopts a more expressive nonlinear activation in the feed-forward network to improve feature learning. Additionally, a Semantic-Pattern Placement Data Augmentation (SPDA) strategy is proposed to increase training data diversity and generalization. Experiments confirm RELSTR outperforms LSTR across multiple metrics, particularly under adverse scenarios, highlighting its potential to enhance safety in intelligent transportation and autonomous driving.

We propose a modular pipeline that converts Chinese text into Taiwanese Hokkien speech and synthesizes a photorealistic talking-face video. The system couples a Taiwanese-aware TTS front end (hybrid word segmentation, iTaigi Tâi-lô romanization with dynamic tone annotation, tone-aware Tacotron2) with a Seed-VC–style timbre converter and a two-stage speech-driven talking-face synthesizer using cross-modal fusion, temporal modeling, and diffusion rendering. On an internal set, the speech reaches MOS 4.602 and CER 2.01%. On talking-face benchmarks, we obtain FID 29.18, CPBD 0.498, and LSE-D 7.23, surpassing representative baselines in realism, sharpness, and lip-sync. The design is lightweight and modular, runs on a single RTX 3080, and is transferable to other low-resource tonal languages.

Traditional animal behavior recognition models require extensive labeled datasets and frequent retraining, limiting their adaptability across species and environments. Additionally, existing systems rely on predefined behavior categories, making it difficult for researchers to customize recognition models to specific behavioral patterns relevant to their studies. Different research fields, such as animal welfare monitoring, conservation, and behavioral ecology, often require distinct behavior classifications, yet current systems lack the flexibility to accommodate these varying needs. This study aims to develop an expandable and user-driven animal behavior recognition system utilizing DeepLabCut for pose estimation and a BiLSTM-based classification model. By integrating Dynamic Adaptive Elastic Weight Consolidation (DA-EWC), the system enables incremental learning, allowing new behaviors to be added with minimal annotation while preserving previously learned behaviors. The proposed system is trained on dolphin behavior datasets using DeepLabCut for keypoint extraction and a BiLSTM model for sequence classification. Additionally, a user-friendly interface is implemented to facilitate behavior annotation and efficient model updates. The proposed system achieves 96.5% accuracy in behavior classification, surpassing conventional models such as Faster R-CNN. Compared to standard EWC, DA-EWC maintains an average of 8.3% higher accuracy when incorporating new behaviors. Furthermore, the system reduces annotation efforts by 9.3%, enabling users to expand behavior categories efficiently. This expandable behavior recognition system significantly enhances adaptability and efficiency in animal behavior monitoring. By supporting user-driven incremental learning, it provides a scalable solution for behavior analysis across different research domains, addressing the need for customizable and evolving behavior classification.

With the growing global focus on sustainability, the demand for ESG reports that meet international guidelines is rising. However, existing systems often face incomplete datasets and unclear methodologies. This study proposes a novel method integrating ChatGPT-4o with an ESG-enhanced hybrid RAG (ESGH-RAG) for ESG report generation, utilizing a high-performance computing (HPC) environment to enhance processing speed. This method introduces two key innovations: first, ESGH-RAG allows for the retrieval of ESG guidelines and dynamically adjusts retrieval methods; second, ChatGPT-4o Report Synthesizer (CRS) generates reports through a multi-prompt module. This method was tested using real-world ESG datasets from publicly listed companies, with experiments conducted on two hardware setups. Results show that in an HPC environment, report generation is faster, and the method outperforms existing approaches in generating high-quality ESG reports. ESGH-RAG achieved 92.1

Fatigue and distracted driving are among the leading causes of traffic accidents, highlighting the importance of developing efficient and non-intrusive driver monitoring systems. Traditional camera-based methods are often limited by lighting variations, occlusions, and privacy concerns. In contrast, millimeter-wave (mmWave) radar offers a non-contact, privacy-preserving, and environment-robust solution, providing a forward-looking alternative. This study introduces a novel deep learning model, RTSFN (radar-based temporal-spatial fusion network), which simultaneously analyzes the temporal motion changes and spatial posture features of the driver. RTSFN incorporates a cross-gated fusion mechanism that dynamically integrates multi-modal information, enhancing feature complementarity and stabilizing behavior recognition. Experimental results show that RTSFN effectively detects dangerous driving states with an average F1 score of 94% and recognizes specific high-risk behaviors with an average F1 score of 97% and can run in real-time on edge devices such as the NVIDIA Jetson Orin Nano, demonstrating its strong potential for deployment in intelligent transportation and in-vehicle safety systems.

In recent years, artificial intelligence (AI) approaches in computer vision and medical technology have been combined to create various convenient and accurate tools to assist medical treatments. In this work, we propose conditional generative adversarial networks (conditional GANs)-based pigmented facial skin analysis system for melasma diagnosis. In the past, melasma diagnosis was based on subjective diagnoses from doctors, and there were few automatic melasma analysis methods. The proposed system helps to determine the region according to the melasma’s severity. Areas associated with melasma and hemoglobin are detected to determine whether they may require special treatments. Furthermore, the proposed work cooperates with HUANGDERM dermatology to collect a facial skin pigmented dataset. We divide the dataset into 3,000 groups for training datasets and 678 groups for testing. Each group contains four categories of images: standard white light, polarized light, melanin and hemoglobin distribution. As a result, the proposed system successfully generates melasma and hemoglobin images and performs well with respect to subjective and objective evaluations.

Caring for dolphins is a delicate process that requires experienced caretakers to pay close attention to their behavioral characteristics. However, caretakers may sometimes lack experience or not be able to give their full attention, which can lead to misjudgment or oversight. To address these issues, a dolphin behavior analysis system has been designed to assist caretakers in making accurate assessments. This study utilized image preprocessing techniques to reduce sunlight reflection in the pool and enhance the outline of dolphins, making it easier to analyze their movements. The dolphins were divided into 11 key points using an open-source tool called DeepLabCut, which accurately helped mark various body parts for skeletal detection. The AquaAI Dolphin Decoder (ADD) was then used to analyze six dolphin behaviors. To improve behavior recognition accuracy, the long short-term memory (LSTM) neural network was introduced. The ADD and LSTM models were integrated to form the ADD-LSTM system. Several classification models, including unidirectional and bidirectional LSTM, GRU, and SVM, were compared. The results showed that the ADD module combined with a double-layer bidirectional LSTM method achieved high accuracy in dolphin behavior analysis. The accuracy rates for each behavior exceeded 90%.

In Taiwan, traditional production equipment for the mainframe panels is imported from overseas, and the parameters are adjusted through the operation panel for automated manufacturing. However, these parameters are slightly different from those used in Taiwan, requiring manual recording of warning parameters to ensure quality control. Currently, the only way to capture the operation records of these machines is by manually recording the system panel information, which is a time-consuming and laborious process that is expensive for production line personnel. In this work, we use image recognition to capture and analyze data externally without damaging the machine. This research improves the ESRGAN network to restore the image details and textures of the mainframe panels. It also captures relevant data with Google OCR to convert panel images into parameter values. The captured results are combined with data analysis to provide more accurate standard mainframe panel information. Even under different interference source conditions, data extraction and analysis can still be carried out to output the text files, effectively assisting the production line in recording the machine panel parameters and reducing personnel's workload.

In AI pandemic applications, the online automatic AI recording apparatus for official councils such as court trials, business conferences and commercial meetings will become imperative because it could let the opinion identification and consensus of participants be synchronically available to implicitly diminish social costs such as follow-up disputes and controversies. Hence, in this study, an automatic on-line multi-dialogue recording system is completely constructed, where the unbounded interleaved-state recurrent neural networks (UIS-RNN) with proposed crux improvements is exploited to achieve confident speaker-diarization. For keeping the systematic robustness, a denoising spectral-LSTM, which is precisely modified from the dual-signal transformation LSTM (DTLN), can strengthen its subsequent crux-improved UIS-RNN and automatic speech recognition (ASR). Finally, the MacBERT model is set to rectify the possible wrong words in conversed sentences according to the learned rational context. For making our system being a practical software apparatus in the use of unmarked multi-person councils, we have also completed the convenient interfaces for the operations of ASR and speaker-diarization, which can exhibit on-line denoising efficacy and speaker-diarization results as well as offer real-time hand-crafted rectifications to common users. In extensive experiments, the proposed recording system can promise high accuracy rates of online speaker diarization and speech-separated ASR. Our proposed system had been examined by the cooperated law court staffs, who offered the noise-embedded speeches of practical court field to test our system. Since the tight recording burden had been indeed noticeably alleviated in their legal-action councils, the court staffs had endorsed that the proposed entire system could be a friendly labor-saving AI apparatus for on-line automatic multi-dialogue recording.

Road segmentation is beneficial to build a vision-controllable mission-oriented self-driving bot, e.g., the Self-Driving Sweeping Bot, or SDSB, for working in restricted areas. Using road segmentation, the bot itself and physical facilities may be protected and the sweeping efficiency of the SDSB promoted. However, roads in the real world are generally exposed to intricate noise conditions as a result of changing weather and climate effects; these include sunshine spots, shadowing caused by trees or physical facilities, traffic obstacles and signs, and cracks or sealing signs resulting from long-term road usage, as well as different types of road materials, such as cement or asphalt; all of these factors greatly influence the effectiveness of road segmentation. In this work, we investigate the extension of Primordial U-Net by the proposed EnRDeA U-Net, which uses an input channel applying a Residual U-Net block as an encoder and an attention gate in the output channel as a decoder, to validate a dataset of intricate road noises. In addition, we carry out a detailed analysis of the nets' features and segmentation performance to validate the intricate noises dataset on three U-Net extensions, i.e., the Primordial U-Net, Residual U-Net, and EnRDeA U-Net. Finally, the nets' structures, parameters, training losses, performance indexes, etc., are presented and discussed in the experimental results.

Abstract Convolutional deep learning models have shown comparable performance to radiologists in detecting and classifying thoracic diseases. However, research on rib fractures remains limited compared to other thoracic abnormalities. Moreover, existing deep learning models primarily focus on using frontal chest X‐ray (CXR) images. To address these gaps, the authors utilised the EDARib‐CXR dataset, comprising 369 frontal and 829 oblique CXRs. These X‐rays were annotated by experienced radiologists, specifically identifying the presence of rib fractures using bounding‐box‐level annotations. The authors introduce two detection models, AB‐YOLOv5 and PB‐YOLOv5, and train and evaluate them on the EDARib‐CXR dataset. AB‐YOLOv5 is a modified YOLOv5 network that incorporates an auxiliary branch to enhance the resolution of feature maps in the final convolutional network layer. On the other hand, PB‐YOLOv5 maintains the same structure as the original YOLOv5 but employs image patches during training to preserve features of small objects in downsampled images. Furthermore, the authors propose a novel two‐level cascaded architecture that integrates both AB‐YOLOv5 and PB‐YOLOv5 detection models. This structure demonstrates improved metrics on the test set, achieving an AP30 score of 0.785. Consequently, the study successfully develops deep learning‐based detectors capable of identifying and localising fractured ribs in both frontal and oblique CXR images.

According to statistics from the Ministry of Transportation, distracted driving is one of the main causes of traffic accidents. In Taiwan, approximately 20% of total accidents each year are attributed to distracted or fatigued driving, ranking it as the second leading cause of accidents. Typically, drivers start to feel fatigued and experience decreased reaction and decision-making abilities after driving continuously for 2-3 hours. Post-meal and midnight driving also tend to induce drowsiness more than other times. Additionally, alcohol consumption, drug use, and the use of hypnotic medication can lead to loss of consciousness. Fatigued driving poses a serious public safety issue, endangering both drivers and other road users. Therefore, developing effective methods for detecting driver fatigue is crucial. Based on existing research, employing facial monitoring systems can reduce accident rates by 10% to 20%. However, these systems are limited by lighting conditions. Thus, this paper proposes the use of millimeter-wave technology to realize a fatigue detection system, effectively addressing the high reliance of facial monitoring systems on lighting conditions. In this study, data collection was performed using the AWR1642boost, a highly integrated millimeter-wave radar development board developed by Texas Instruments. Subsequently, a Multi-experts system was developed to integrate and analyze four features: nodding, yawning, breathing, and heart rate. The system employed a Support Vector Machine (SVM) model to improve the accuracy of fatigue detection in driving scenarios. Experimental results demonstrated that the system achieved an accuracy rate of 90%, effectively discerning whether the driver was experiencing fatigue.

With the increasing popularity of artificial intelligence, the technologies used by robots are changing rapidly. We propose a robotic system that can handle Mandarin-related problems, effectively dealing with Mandarin phonetic transcription, multiple-option reading comprehension, text error correction, similarity judgment, and keyword search. The experimental results show that using the national language assessment of grades 1–6 in primary schools, the highest score on the national language assessment is 90, and the lowest is 63, which is consistent with the actual score-provided by Shekou National Primary School for Mandarin assessment. Furthermore, more than half of the average scores in each subject are higher than the low standard (average of the bottom 25%) and even higher than the high standard (average of the top 25%). Through this system, the machine can reach the national language level of primary school students.

In recent years, the research of autonomous robotic arms has received high attention from academia and industry, so this thesis aims to develop a visual detection system for robotic arms to grasp and place objects. Object detection technology The purpose of which is to locate objects of interest (target) in a given image, providing object location bounding boxes and categories. ROS is a set of open-source software libraries designed to simplify the creation of complex and robust on various robotic platforms. the task of the robot behavior. The proposed scheme uses a camera with depth information and combines target localization and edge detection algorithms to accurately measure the relative distance between the object and the robot arm. The process includes deep neural network construction and use, robotic arm positioning, object bounding box position information, and robotic arm movement control. Considering the visual scale, the relative distance between the target and the robotic arm is calculated so that the robotic arm can grasp the target and place the object in a specific position. This paper introduces how to combine the deep learning object detection model and ROS moveit! to complete the object identification and grasp and place the robotic arm.

The Self-Driving Sweeper Bot (SDSB) was previously reported by incorporated the Self-Driving system with the Sweeper mechanism, after several months and numerous times of running test at a pre-planned close-loop campus route, our observation with the best knowledge indicates that SDSB is extendable to be SDX tension B, such as SD LM B incorporated Self-Driving Bot with Lawn Mower mechanism, SDPB integrated Self-Driving Bot with Paver mechanism and SD SK B is together the Self-Driving Bot with StacKer etc. To reach SDX tension B with visual intelligence and human knowledge capabilities, the observation further reports that three important issues have to be further inspecting, including, 1) complex environmental noises for machine visual learning, 2) modelling trigger-event for failed event to functionality restoration, and 3) sensors security against hacker remote control. Besides that, the investigation on SDSB visual intelligence in terms of road segmentation, rubbish, pedestrian and vehicle detections by YOLO and U-net extension approaches were validated on its promising accuracy and feasibility to reach SDSB visual intelligence.