Xuyi Yu

Xuyi Yu

Auto-generated profile, based on publicly available information

Xuyi Yu

State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xian Jiaotong University, Xian, China

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 April 2022 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
XYXuyi Yu
Xuyi Yu
WLWentao Lyu
Wentao Lyu
DZDi Zhou
Di Zhou
CWChengqun Wang
Chengqun Wang
WXWeiqiang Xu
Weiqiang Xu
Defect detection is to locate and classify the possible defects in an image, which plays a key role in the quality inspection link in the manufacturing process of industrial products. Defects in industrial products are generally very small and extremely uneven in scale, resulting in poor detection results. Therefore, we propose an efficient scale-aware network (ES-Net) to improve the effect of defect detection. By addressing the information loss of tiny targets and the mismatch between the receptive field of detection head and the scale of targets, ES-Net improves the overall defect detection effect, especially for tiny defects. Considering that existing works directly use an integrated feature to enhance features at all levels, it may cause confusion in the direction of network optimization. Therefore, we propose the aggregated feature guidance module (AFGM), which first performs efficient cascading fusion of multi-level features to filter cross-layer conflicts. Then the split and aggregation enhancement (SAE) module is designed to further optimize the integrated feature map, and the result is used to guide the shallow features. Moreover, we also introduce the multi-receptive field fusion (MFF) module to generate multi-receptive field information to supplement the shallow features after dimensionality reduction. The efficient stair pyramid (ESP) is a further improvement of feature pyramid network (FPN)-based network. In particular, we propose the dynamic scale-aware head (DSH) in shallow detection layer, which can adaptively select the best detection receptive field according to different scales of targets, thereby improving the detection performance of tiny targets. Extensive experimental results on Aliyun Tianchi fabric dataset (76.2% mAP), NEU-DET (79.1% mAP), and printed circuit board (PCB) defect dataset of Peking University (97.5% mAP) demonstrate the proposed ES-Net achieves competitive results compared to the state-of-the-art (SOTA) methods. Moreover, the high efficiency of ES-Net makes it more applicable in scenarios with high real-time requirements.

30 January 2026 in Neural Networks
Verified
Journal Article
Closed access
XYXuyi Yu
Xuyi Yu
SLShang Lou
Shang Lou
YZYinghai Zhao
Yinghai Zhao
HZHuipeng Zhang
Huipeng Zhang
KMKuizhi Mei
Kuizhi Mei

7 November 2025 in Measurement Science and Technology
Verified
Journal Article
Closed access
ZSZiming Song
Ziming Song
XYXuyi Yu
Xuyi Yu
YLYanchun Liang
Yanchun Liang
ZYZicong Yang
Zicong Yang
Due to the high complexity and technical requirements of industrial production processes, surface defects will inevitably appear, which seriously affect the quality of products. Although existing lightweight detection networks are highly efficient, they are susceptible to false or missed detection of non-salient defects due to the lack of semantic information. In contrast, the diffusion model can generate higher-order semantic representations in the denoising process. Therefore, this paper aims to incorporate the higher-order modeling capability of diffusion models into the detection framework, to better support the classification and localization of challenging targets. First, the denoising diffusion probabilistic model (DDPM) is pre-trained to extract the features of the denoising process to construct a feature repository. In particular, to avoid the potential bottleneck of memory caused by the dataloader loading high-dimensional features, a residual convolutional variational auto-encoder is designed to further compress the feature repository. The image is fed into both the image backbone and feature repository for feature extraction and querying respectively. The queried latent features are reconstructed and filtered to obtain high-dimensional DDPM features. A dynamic cross-fusion method is proposed to fully refine the contextual features of DDPM to optimize the detection model. Finally, we employ knowledge distillation to migrate the higher-order modeling capabilities back into the lightweight baseline model without additional efficiency cost. Experiment results demonstrate that our method achieves competitive results on several industrial datasets.

1 April 2026 in Machine Vision and Applications
Verified
Journal Article
Closed access
PDF file
HZHuipeng Zhang
Huipeng Zhang
XYXuyi Yu
Xuyi Yu
KMKuizhi Mei
Kuizhi Mei

1 January 2026 in IEEE Transactions on Circuits and Systems for Video Technology
Verified
Journal Article
Closed access
HZHuipeng Zhang
Huipeng Zhang
XYXuyi Yu
Xuyi Yu
MJMiaomiao Jin
Miaomiao Jin
KMKuizhi Mei
Kuizhi Mei

29 April 2026 in IEEE Transactions on Circuits and Systems for Video Technology
Verified
Journal Article
Closed access
XYXuyi Yu
Xuyi Yu
YLYaohua Liu
Yaohua Liu
YZYinghai Zhao
Yinghai Zhao
HZHuipeng Zhang
Huipeng Zhang
KMKuizhi Mei
Kuizhi Mei
Efficient model compression is essential for deploying semantic segmentation models on edge devices. While quantization-aware training combined with knowledge distillation (KD) has emerged as a dominant paradigm, existing methods face a critical challenge: quantized teacher features provide deployment-compatible supervision but suffer from information loss, while full-precision features retain rich semantic details but exhibit representation mismatch with quantized students. Our analysis reveals that while single supervision sources each have inherent limitations, their combination produces superior results, suggesting the existence of more favorable intermediate representations in the feature space. To systematically explore this advantageous representation space, we propose Multi-perspective Feature Distillation (MFD). Our framework constructs multiple virtual quantized teachers by strategically perturbing the quantization parameters of a single full-precision teacher, generating diverse supervisory signals that bridge the gap between full-precision and quantized representations. This multi-perspective approach guides the student model toward learning robust features that maintain semantic richness while accommodating quantization constraints, all with minimal computational overhead. Comprehensive experiments show that MFD brings substantial improvements over existing distillation methods under 8-bit settings, while being efficient to train and readily integrable with existing KD frameworks. The source code will be available at https://github.com/Fateeeeee/MFD.