ZZZhiying Zhou

Zhiying Zhou

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zhiying Zhou has a total of 37 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Enhanced Radar Fall Detection via Compressed Sensing and Deep Learning-Assisted Point Cloud Generation

    by Zhi Zheng, Siyuan Zhao, Bo Wang, Zhiying Zhou, Xudong Chen, Yongxin Guo

    Existing fall detection based on radar point cloud faces the challenges of low resolution point cloud estimation and sparsity, which seriously restricts the system detection performance. This study proposes an innovative radar point cloud generation method combining compressed sensing and a depth camera to improve the accuracy and robustness of the fall detection system. Specifically, the study reformulates radar point cloud direction of arrival (DOA) estimation as a linear inverse problem and solves it using compressed sensing (CS) with the fast approximate message passing (AMP) algorithm, significantly improving point cloud positional resolution compared with the traditional fast Fourier transform (FFT) method. Subsequently, an Encoder-Decoder network based on the attention mechanism is designed to effectively enhance the density and expression ability of the point cloud by capturing the local features of the radar point cloud. The depth camera in this article is only used in the network training process. In the actual fall detection stage, the system only uses radar signals for classification. The experimental results verify the advantages of the proposed method in indoor fall detection. Compared with the traditional FFT-based point cloud generation results, the detection accuracy is improved from ∼ 91% to ∼ 96%. The innovative method proposed in this paper provides an efficient, economical, and easy-to-deploy technical solution for fall detection systems in the smart home environment, which has significant potential applications.

  • Open Access

    A Decision-support Framework for Energy-efficient Batch Optimization in Machine Learning

    by Timothy A. Adeyi, Adrian D. Cheok, Huihui Song, Zhiying Y. Zhou, Emma Y. Zhang, Wisura P. Pallewatta, Ishmeal Quist

    This paper presents a decision-support framework for selecting energy-efficient batch optimization techniques in machine learning, addressing the absence of reproducible, theory-informed guidance in Green Artificial Intelligence (AI). Eight key methods, including Mixed Precision Training, Dynamic Batching, and Gradient Accumulation are systematically evaluated through a rigorous systematic literature review and assessed across five criteria: energy efficiency, accuracy impact, scalability, technological maturity, and hardware dependency. We operationalize a transparent, weighted Multi-Criteria Decision Analysis (MCDA) protocol, termed the Green AI Score, and a normalized efficiency metric, Energy-Delay Normalized Accuracy (EDNA). These instruments serve as a reproducible analytical scaffold to enable cross-technique comparison and quantification of energy-accuracy trade-offs. Advanced visual analytics, including a heatmap, dual-axis bar chart, bubble plot, and correlation visualization, link sustainability rankings with energy-accuracy trade-offs. Our analysis identifies Mixed Precision Training (Score: 0.84) and Dynamic Batching (Score: 0.77) as top-tier strategies, offering 40–60% and 30–50% energy savings, respectively, with minimal accuracy loss. Furthermore, we theoretically derive a hardware-aware heuristic threshold (memory-bandwidth-to-compute ratio < 0.4) to analytically assess the applicability of Mixed Precision Training under varying system constraints. Finally, a rule-based decision policy is synthesized to support informed technique selection, providing practitioners with a theoretically grounded guide for navigating deployment trade-offs. This work advances beyond descriptive surveying by providing an evidence-based, reproducible analytical synthesis grounded in systems and algorithmic considerations, thereby supporting transparent and sustainable AI deployment.

  • Open Access

    Adaptively Distilled ControlNet: Accelerated Training and Superior Sampling for Medical Image Synthesis

    by Kunpeng Qiu, Zhiying Zhou, Yongxin Guo

    Medical image annotation is constrained by privacy concerns and labor-intensive labeling, significantly limiting the performance and generalization of segmentation models. While mask-controllable diffusion models excel in synthesis, they struggle with precise lesion-mask alignment. We propose \textbf{Adaptively Distilled ControlNet}, a task-agnostic framework that accelerates training and optimization through dual-model distillation. Specifically, during training, a teacher model, conditioned on mask-image pairs, regularizes a mask-only student model via predicted noise alignment in parameter space, further enhanced by adaptive regularization based on lesion-background ratios. During sampling, only the student model is used, enabling privacy-preserving medical image generation. Comprehensive evaluations on two distinct medical datasets demonstrate state-of-the-art performance: TransUNet improves mDice/mIoU by 2.4%/4.2% on KiTS19, while SANet achieves 2.6%/3.5% gains on Polyps, highlighting its effectiveness and superiority. Code is available at GitHub.

  • Efficient needle guidance: multi-camera augmented reality navigation without patient-specific calibration

    by Yizhi Wei, Bingyu Huang, Bolin Zhao, Zhengyu Lin, Steven Zhiying Zhou

    Purpose Augmented reality (AR) technology holds significant promise for enhancing surgical navigation in needle-based procedures such as biopsies and ablations. However, most existing AR systems rely on patient-specific markers, which disrupt clinical workflows and require time-consuming preoperative calibrations, thereby hindering operational efficiency and precision. Methods We developed a novel multi-camera AR navigation system that eliminates the need for patient-specific markers by utilizing ceiling-mounted markers mapped to fixed medical imaging devices. A hierarchical optimization framework integrates both marker mapping and multi-camera calibration. Deep learning techniques are employed to enhance marker detection and registration accuracy. Additionally, a vision-based pose compensation method is implemented to mitigate errors caused by patient movement, improving overall positional accuracy. Results Validation through phantom experiments and simulated clinical scenarios demonstrated an average puncture accuracy of 3.72 ± 1.21 mm. The system reduced needle placement time by 20 s compared to traditional marker-based methods. It also effectively corrected errors induced by patient movement, with a mean positional error of 0.38 pixels and an angular deviation of 0.51\(^\circ \). These results highlight the system’s precision, adaptability, and reliability in realistic surgical conditions. Conclusion This marker-free AR guidance system significantly streamlines surgical workflows while enhancing needle navigation accuracy. Its simplicity, cost-effectiveness, and adaptability make it an ideal solution for both high- and low-resource clinical environments, offering the potential for improved precision, reduced procedural time, and better patient outcomes.

  • Open Access

    Noise-Consistent Siamese-Diffusion for Medical Image Synthesis and Segmentation

    by Kunpeng Qiu, Zhiqiang Gao, Zhiying Zhou, Mingjie Sun, Yongxin Guo

    Deep learning has revolutionized medical image segmentation, yet its full potential remains constrained by the paucity of annotated datasets. While diffusion models have emerged as a promising approach for generating synthetic image-mask pairs to augment these datasets, they paradoxically suffer from the same data scarcity challenges they aim to mitigate. Traditional mask-only models frequently yield low-fidelity images due to their inability to adequately capture morphological intricacies, which can critically compromise the robustness and reliability of segmentation models. To alleviate this limitation, we introduce Siamese-Diffusion, a novel dual-component model comprising Mask-Diffusion and Image-Diffusion. During training, a Noise Consistency Loss is introduced between these components to enhance the morphological fidelity of Mask-Diffusion in the parameter space. During sampling, only Mask-Diffusion is used, ensuring diversity and scalability. Comprehensive experiments demonstrate the superiority of our method. Siamese-Diffusion boosts SANet's mDice and mIoU by 3.6% and 4.4% on the Polyps, while UNet improves by 1.52% and 1.64% on the ISIC2018. Code is available at GitHub.

  • Contactless High Dynamic-Range Blood Pressure Estimation by Pulse Morphological Features Based on Accurate Doppler Radar Detection

    by Zhi Zheng, Shuqin Dong, Haotian Shi, Bo Wang, Zhiying Zhou, Yongxin Guo

    Blood pressure (BP) can be estimated contactlessly based on Doppler wrist pulse (DWP) detection. However, this technique faces the challenge of interferences from the respiratory and its harmonic components, which can limit the accuracy of the BP estimation. This article first presents an adaptive respiratory harmonic filtering (ARHF) technique, aiming to mitigate the interference of respiration on the wrist pulse signal. Also, a customized 24-GHz continuous wave radar system with high linearity and a high gain antenna is used for better DWP detection. Second, a novel radar-based BP estimation model based on multidimensional morphological features extraction is proposed for high-accuracy and high-robustness BP estimation. In a stationary BP estimation experiment with eight subjects over three consecutive days, the proposed method demonstrates an average mean error and standard deviation of 2.60 $\pm$ 7.89 mmHg and 1.01 $\pm$ 2.84 mmHg in the sitting posture for systolic and diastolic blood pressure (SBP and DBP), respectively, achieving the highest estimation accuracy among the recent contactless radar-based BP detection. Additionally, a specific BP fluctuation range experiment is conducted to continuously measure BP, including varying BP fluctuations across five stages. In the worst case scenario, the proposed method exhibits SBP and DBP errors of 2.50 $\pm$ 2.98 and 8.34 $\pm$ 9.86 mmHg, respectively, affirming the robustness of the proposed technique. The experimental findings highlight the feasibility and reliability of the DWP extraction and subsequent contactless BP estimation techniques in the healthcare domain.


  • Open Access

    Investigating the accuracy of machine vision and augmented reality in percutaneous computed tomography-guided interventions: A phantom study

    by Bingyu Huang, Yizhi Wei, Bing Zhang, Jin Chen, Rui Guo, Steven Zhiying Zhou, Zhigang Lin, Zhengyu Lin

    Objectives: This study aimed to evaluate the accuracy of percutaneous computed tomography (CT)-guided puncture based on machine vision and augmented reality in a phantom.Materials and Methods: The surgical space coordinate system was established, and accurate registration was ensured using the hierarchical optimization framework. Machine vision tracking and augmented reality display technologies were used for puncture navigation. CT was performed on a phantom, and puncture paths with three different lengths were planned from the surface of the phantom to the metal ball. Puncture accuracy was evaluated by measuring the target positioning error (TPE), lateral error (LE), angular error (AE), and first success rate (FSR) based on the obtained CT images.Results: A highly qualified attending interventional physician performed a total of 30 punctures using puncture navigation. For the short distance (4.5–5.5 cm), the TPE, LE, AE, and FSR were 1.90 ± 0.62 mm, 1.23 ± 0.70 mm, 1.39 ± 0.86°, and 60%, respectively. For the medium distance (9.5–10.5 cm), the TPE, LE, AE, and FSR were 2.35 ± 0.95 mm, 2.00 ± 1.07 mm, 1.20 ± 0.62°, and 40%, respectively. For the long distance (14.5–15.5 cm), the TPE, LE, AE, and FSR were 2.81 ± 1.17 mm, 2.33 ± 1.34 mm, 0.99 ± 0.55°, and 30%, respectively.Conclusion: The augmented reality and machine vision-based CT-guided puncture navigation system allows for precise punctures in a phantom. Further studies are needed to explore its clinical applicability.


  • Radar-Based Short-Time Frequency Accumulation for Masked Fall Detection

    by Zhi Zheng, Bo Wang, Zhiying Zhou, Yongxin Guo

    Masked situations, prevalent in complex scenarios involving multiple people or objects, can be defined as the case where the line of sight between radar and a target is partially obstructed. These situations, with degraded information received by radar, bring challenges for fall detection using conventional approaches. This paper introduces a novel approach to detecting falls in masked scenarios using a millimeter-wave (mmWave) radar. The proposed approach combines two main features extracted from the radar signals: a) the Pattern-Contour-Confined DC-Corrected Doppler time (PCC-DCDT) map, generated from DC-corrected radar signals; b) the Short-Time Frequency Accumulation (STFA), proposed for enhancement of the weakened information in masked situations. Experimental results demonstrate that the combination of the PCC-DCDT map and STFA improves the precision and reliability in masked situations. This exploration work on masked fall scenarios adapts radar sensors to realistic indoor environments with improved accuracy, providing sufficient detection robustness for ubiquitous healthcare applications.


  • Contrastive Attention for Video Anomaly Detection

    by Shuning Chang, Yanchao Li, Shengmei Shen, Jiashi Feng, Zhiying Zhou

    We consider weakly-supervised video anomaly detection in this work. This task aims to learn to localize video frames containing anomaly events with only binary video-level annotation, i.e. , anomaly vs. normal. Traditional approaches usually formulate it as a multiple instance learning problem, which ignore the intrinsic data imbalance issue that positive samples are very scarce compared to negative ones. In this paper, we focus on addressing this issue to boost detection performance further. We develop a new light-weight anomaly detection model that fully utilizes enough normal videos to train a classifier with a good discriminative ability for normal videos, and we employ it to improve the selectivity for anomalous segments and filter out normal segments. Specifically, in addition to boosting anomalous prediction, a novel contrastive attention module additionally produces a converted normal feature from anomalous video to refined anomalous predictions by maximizing the classifier making a mistake. Moreover, to remove the stubborn normal segments selected by the attention module, we also design an attention consistency loss to employ the classifier with high confidence for normal features to guide the attention module. Extensive experiments on two large-scale datasets, UCF-Crime, ShanghaiTech and XD-Violence, clearly demonstrate that our model largely improves frame-level AUC over the state-of-the-art. Code is released at https://github.com/changsn/Contrastive-Attention-for-Video-Anomaly-Detection .