PWPeishu Wu

Peishu Wu

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Peishu Wu has a total of 61 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • A Tensor-Attention and DARTS-Guided Detection Framework for Aeroengine Blades’ Surface Defects

    by Xiao Wang, Peishu Wu, Da Teng, Haonan Chen, Xiaoyang Bi, Nianyin Zeng

    The reliable, real-time detection of surface defects on aero-engine blades is severely limited by complex geometry and lighting variations, which reduce detection precision under strict computational budgets. To address these coupled challenges, we present a lightweight detection framework, termed Tensor-Attention- and DARTS-Guided Network (TAD-Net), featuring a task-tailored architectural design that improves defect detectability while maintaining deployment-level efficiency for aero-engine inspection. TAD-Net consists of three synergistic components. (i) A Tensor-Attention Fusion Block enhances long-range dependency modeling for weak defect cues while suppressing redundant shallow activations. (ii) A DARTS-guided Residual Attentive Large-Receptive-Field Neck improves multi-scale feature fusion by enforcing cross-scale alignment and representation consistency. (iii) A Hierarchical-Rank Tensor Decomposition strategy reduces high-dimensional feature redundancy with minimal loss of semantic capacity, thereby lowering computational cost. These components are coordinated within a unified design to address the interplay between feature representation, multi-scale aggregation, and computational efficiency in aeroengine inspection scenarios. TAD-Net is evaluated on four benchmark datasets and consistently improves the accuracy and efficiency balance. On the Aeroengine-Blade dataset, average precision increases by about 4%, while computational cost can be reduced by over 10% through the proposed design, with comparable recall. These gains support real-time deployment in resource-limited aero-engine inspection systems, demonstrating the suitability of TAD-Net for integration into sensor-driven inspection and monitoring pipelines in safety-critical industrial applications.


  • DartsNeXt: Bridging the Search-Evaluation Gap in Differentiable NAS With Router-Based Selection and Sequential Architecture

    by Liwei Hu, Zidong Wang, Peishu Wu, Keyi Yu, Nianyin Zeng

    Neural Architecture Search (NAS) has emerged as a pivotal approach for automating the design of high-performance neural networks, with Differentiable Architecture Search (DARTS) widely being recognized for its efficiency. However, DARTS suffers from critical limitations, including a soft–hard mismatch and a depth mismatch between the search and evaluation phases, leading to instability and performance collapse. To address these issues, we propose DartsNeXt, a sequential differentiable NAS framework that aligns search and evaluation across both the operator and macro levels. At the operator level, we introduce a novel router-based selection method that replaces magnitude-based selection with input-conditioned sparse routing, ensuring that the forward graph during search is consistent with the final discrete architecture while reducing bias toward skip connections. At the macro level, a stage-wise backbone with large-kernel depthwise convolutions minimizes memory usage and mitigates the depth discrepancy, enabling search to be performed directly at the evaluation depth. Extensive experiments demonstrate that DartsNeXt achieves stable and robust performance on NAS-Bench-201 and the DARTS search space. Furthermore, DartsNeXt provides a fully automated and compute-efficient pipeline that discovers competitive token mixers without manual design, delivering near-expert performance on ImageNet-1 K classification and COCO detection tasks under realistic compute budgets.


  • A novel microarray detection method based on deep learning with edge computing

    by Zhenhua Gan, Dongyu He, Peishu Wu, Baoping Xiong, Nianyin Zeng, Lyuchao Liao, Feng Guo, Qin Bao

    Traditional microarray scanners typically rely on high-performance computers to process the acquired images. To address these challenges, we propose a novel microarray detection (MD) method designed for deployment on a charge-coupled device (CCD) microarray scanner integrated with an edge computing. In this MD method, we propose a novel model named C2f_SC, which integrates the computational efficiency of the star operation with contextual anchor attention to enhance detection accuracy within a compact model architecture. We further integrate the C2f_SC module into the You Only Look Once v8 (YOLOv8) framework, utilizing MobileNetV3 with the convolutional block attention module (CBAM) as the backbone, removing the detection heads for medium and large objects, replacing the bounding box loss function with the signed intersection over union, and incorporating ResBlock_CBAM before the small object detection head. With the microarray images from the gene expression omnibus database, experimental results demonstrate that our method significantly reduces the required giga floating point operations per seconds (GFLOPs) and model size compared to other YOLO models, while maintaining comparable or higher mAP@50 performance with only a slight drop at mAP@50–95. Specifically, against YOLOv8n, our approach matches mAP@50 while using just 32.9% of GFLOPs and 43.3% of model weights. In addition, we conducted a comparative experiment between the laser confocal scanner and the CCD scanner on the BAC Microarray from MGmed Inc., Republic of Korea (MGMED BAC) to verify the reliability of the CCD scanner. Extensive experimental results demonstrate that the proposed MD method significantly reduces reliance on large-scale computing systems.

  • A Low-Complexity YOLOv8s-Based Model for Environment Perception in Power Maintenance Guided Vehicles

    by Zhenhua Gan, Jinyang Li, Peishu Wu, Dongyu He, Lyuchao Liao, Nianyin Zeng, Baoping Xiong, Feng Guo, Yuankun Bai

    Traditional methods for moving power covers in the power maintenance sector introduce complexities and safety risks due to their heavy reliance on manual labor. In order to reduce dependence on human resources while enhancing safety in operational processes, we introduce a low-complexity model called YOLO-EPGV (Environmental Perception in Guided Vehicles YOLO), which is built upon YOLOv8s to address this challenge. This model is specifically tailored for environmental perception in engineering vehicles used for power maintenance, aiming to advance vehicle automation, and minimize labor requirements and accident risks. YOLO-EPGV combines the StartNet architecture, Partial Convolution (PConv), and Convolutional Gated Linear Units (CGLU) to improve detection precision (mAP@50) while reduce computational costs. Furthermore, we developed a new greenbelt dataset, which was used for training and evaluating the model. The results show that YOLO-EPGV delivers excellent performance, achieving a mAP@50 of 92.8% with minimal computational load, outperforming several YOLO variants. Ablation studies confirm the contribution of each component to the overall performance, and comparisons with other lightweight backbone networks highlight the superior efficiency and precision of StartNet. Subsequently, the model was deployed on the NVIDIA Jetson Orin Nano, achieving a detection speed of 57 FPS while maintaining mAP@50 of 92.7%. This achievement has been validated through empirical testing in real-world operational settings.

  • From data analysis to intelligent maintenance: a survey on visual defect detection in aero-engines

    by Peishu Wu, Han Li, Xin Luo, Liwei Hu, Rui Yang, Nianyin Zeng

    In this paper, a systematic review of aero-engine defect detection methods is presented, encompassing the general procedure, traditional and intelligent detection algorithms, performance optimization, and future trends. The complete process and innovative theories of aero-engine visual defect detection are analyzed in this overview. Specifically, a five-level taxonomy is designed, with each level further subdivided to provide deeper insights, from data acquisition and task-oriented detection with nondestructive testing (NDT), to practical applications. By leveraging multiscale feature fusion-based detection, these methods achieve enhanced precision in identifying defects across varying scales and complexities. Moreover, in-depth discussions and outlooks on performance optimization and efficient deployment strategies are provided to promote advanced intelligent maintenance solutions for high-end equipment, which may encourage more multidisciplinary collaborations. Compared to other existing surveys, this work comprehensively outlines how computer vision (CV)-based methods can assist in aero-engine defect detection for intelligent decision-making, and a connection between NDT technology and CV-based inspection has been established, thereby drawing greater attention to the application of artificial intelligence to further enhance the development of industrial predictive maintenance.


  • DLA-Net: A Dynamically Learnable Attention Network for Intelligent Surface Visual Inspection of Aero-Engine Blades

    by Haonan Chen, Peishu Wu, Weimin Wen, Nianyin Zeng

    The maintenance of aero-engines is critical for ensuring the safe and reliable operation of aircraft. Given the complexity and variability of defect detection tasks in aero-engine blades, such as varying target scales, irregular spatial distributions, and environmental noise interference, this paper proposes a Dynamically Learnable Attention Network (DLA-Net) to address these challenges. Specifically, in DLA-Net, a learnable branch multi-domain attention (LBMA) mechanism is first developed in multi-scale feature extraction and regional focusing strategy, to adaptively adjust important weights for each task-aware channel, spatial and position. Moreover, a task-aligned dynamic detection head (TADDH) is designed, which improves the classification and localization performance by fusing task-interaction features with multi-scale receptive fields. The layer-wise and multipath coordinate attention mechanisms are integrated into the TADDH, dynamically focusing on more informative layers and enhancing flexibility. In addition, a lightweight convolutional technique has been introduced to alleviate the model complexity. Finally, an enhanced focal powerful Intersection over Union (F-PIoU) loss is employed in the bounding box regression, where a non-monotonic focusing mechanism is incorporated, which accelerates the convergence rate and balances the precision loss that may arise from the simplification of the model. Experimental results demonstrate the effectiveness of DLA-Net compared to other state-of-the-art object detection algorithms, and its compact structure indicates that the performance advantages for real-world aero-engine defect detection tasks on resource-constrained devices.


  • MSSFN: A multi-scale sequence fusion network for CT-based diagnosis of pulmonary complications

    by Hongfu Zeng, Xinyu Li, Haipeng Xu, Keyi Yu, Huihua Hu, Peishu Wu, Nianyin Zeng


  • A Novel Equipment Based on Improved YOLOv5s for Automated Electrical Cover Handling

    by Zhenhua Gan, Jinyang Li, Peishu Wu, Yuankun Bai, Baoping Xiong, Nianyin Zeng, Fumin Zou, Dongyu He, Wenshu Ni

    With the rapid development of the power industry, the number of power cables has increased significantly, making the improvement of maintenance efficiency a critical issue. Reducing accidents caused by the manual handling of heavy electrical cover plates has become a pressing concern. However, the maintenance of power cables often requires precise positioning and efficient movement of these heavy cover plates in energy-poor environments. Existing object detection models typically not only require high-performance computing but also have with high cost and large volume, rendering them quite unsuitable for deployment in the resource-constrained system. To address these challenges, we propose an automated solution that integrates negative pressure adsorption and computer vision techniques. The core module of this method is Cover Plates Edge Line-YOLO (CPEL-YOLO), a lightweight model based on You Only Look Once (YOLO). In CPEL-YOLO, it leverages MobileNetV3 with efficient multiscale attention (EMA) and C3-Faster with EMA to focus the model on precise edge lines with reduced computational complexity. Additionally, Scylla-IoU (SIoU) is incorporated to optimize the training process. A dataset of 1009 cover plate images was constructed for training and evaluation. Experimental results show that CPEL-YOLO achieves a mAP@0.5 of 88.1%, comparable to YOLOv5s, while reducing model size by 42% and improving precision by 4.4%.