Hufei Zhu

Hufei Zhu

Auto-generated profile, based on publicly available information

Hufei Zhu

School of Electronics and Information Engineering, Wuyi University, Jiangmen, China

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

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

Representative Publications

View All Publications
22 April 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Verified
Journal Article
Open access
ZYZilu Ying
Zilu Ying
WKWenyu Ke
Wenyu Ke
YZYikui Zhai
Yikui Zhai
ZLZhihao Long
Zhihao Long
JZJianhong Zhou
Jianhong Zhou
CCC.L. Philip Chen
C.L. Philip Chen
In recent years, deep learning-based Synthetic Aperture Radar (SAR) image detection, recognition, and segmentation models achieve remarkable accuracy when trained on large amounts of SAR image samples. However, the acquisition of SAR images tends to be costly. A practical approach to address this issue involves the use of artificially generated training samples to supplement the dataset. Current SAR image generation methods based on Generative Adversarial Networks have issues with generation fidelity and are challenge to converge. Moreover, most of these methods only consider the spatial domain of SAR images without exploring the characteristics of their frequency domain. To tackle the aforementioned problems, we proposed a lightweight frequency domain-aware SAR image generation model based on the Denoising Diffusion Probabilistic Model. The proposed generative model is capable of producing highly realistic artificial SAR image samples while converging stably. Meanwhile, our research reveals that high-frequency components of SAR images play a crucial role in the generation process, and a frequency adjustment module based on the Fast Fourier Transform is implemented in our generation model to prevent interference of high-frequency components by low-frequency components. Tests on the MSTAR dataset highlight our model's advantages in both generation quality and parameter efficiency compared to other cutting-edge image generation models. An additional ablation study also confirms the effectiveness of our proposed frequency adjustment module, which reduces the Fréchet Inception Distance by 2.85%, confirming our hypothesis that high-frequency components are more crucial to SAR image generation. Code is available at https://github.com/yikuizhai/DiffuSAR.

2 June 2025 in IEEE Transactions on Geoscience and Remote Sensing
Verified
Journal Article
Closed access
ZYZilu Ying
Zilu Ying
YZYijun Zhou
Yijun Zhou
YZYikui Zhai
Yikui Zhai
HZHufei Zhu
Hufei Zhu
Hongsheng ZhangHongsheng Zhang
Hongsheng Zhang
CCC.L. Philip Chen
C.L. Philip Chen
With the rapid advancements in deep learning technology, the field of remote sensing change detection (RSCD) has witnessed significant improvements and innovations. In this context, bitemporal image processing, using features directly extracted by the backbone for subsequent fusion operations, may be obstructed by external environmental factors, potentially limiting the effective capture of complex feature variations. Moreover, overlooking local features during the fusion of bitemporal features can significantly affect the final detection results. As a result, achieving accurate change detection (CD) still encounters various challenges. To tackle these issues, this paper proposes a CD network (AEGL-Net) with Adaptive Multiscale Enhancement (AME) and Global-Local Feature Fusion (GLFF) modules. First, AME enhances features at each stage of backbone extraction through an adaptive strategy, balancing the enhancement of semantic information and texture details. Then, GLFF is used to fuse the bitemporal image features, which enhances the modeling of global dependencies while also fusing shared and context-aware weights to enhance the local features. Finally, the merged features are fed into the decoder to generate precise change maps. Experiments conducted with four open RSCD datasets (LEVIR-CD, S2Looking, SYSU-CD, and UAV-CD) demonstrate that our proposed AEGL-Net outperforms ten state-of-the-art models in the RSCD field. Our code is available at https://github.com/yikuizhai/AEGL-Net.

14 September 2020 in Electronics
Verified
Journal Article
Open access
PDF file
HZHufei Zhu
Hufei Zhu
Wen ChenWen Chen
Wen Chen
YWYanpeng Wu
Yanpeng Wu
Based on the efficient inverse Cholesky factorization, we propose an implementation of OMP (called as version 0, i.e., v0) and its four memory-saving versions (i.e., the proposed v1, v2, v3 and v4). In the simulations, the proposed five versions and the existing OMP implementations have nearly the same numerical errors. Among all the OMP implementations, the proposed v0 needs the least computational complexity, and is the fastest in the simulations for almost all problem sizes. As a tradeoff between computational complexities/time and memory requirements, the proposed v1 seems to be better than all the existing ones when only considering the efficient OMP implementations storing G (i.e., the Gram matrix of the dictionary), the proposed v2 and v3 seem to be better than the only existing one when only considering the efficient implementations not storing G, and the proposed v4 seems to be better than the naive implementation that has the (known) minimum memory requirements. Moreover, all the proposed five versions only include parallelizable matrix-vector products in each iteration, and do not need any back-substitutions that are necessary in some existing efficient implementations (e.g., those utilizing the Cholesky factorization).

18 January 2021 in IEEE Access
Verified
Journal Article
Open access
HZHufei Zhu
Hufei Zhu
ZLZhulin Liu
Zhulin Liu
CCC. L. Philip Chen
C. L. Philip Chen
YLYanyang Liang
Yanyang Liang
In this paper, we propose an efficient algorithm to accelerate the existing Broad Learning System (BLS) algorithm for new added nodes. The existing BLS algorithm computes the output weights from the pseudoinverse with the ridge regression approximation, and updates the pseudoinverse iteratively. As a comparison, the proposed BLS algorithm computes the output weights from the inverse Cholesky factor of the Hermitian matrix in the calculation of the pseudoinverse, and updates the inverse Cholesky factor efficiently. Since the Hermitian matrix in the definition of the pseudoinverse is smaller than the pseudoinverse, the proposed BLS algorithm can reduce the computational complexity, and usually requires less than $\frac {2}{3}$ of complexities with respect to the existing BLS algorithm. Our experiments on the Modified National Institute of Standards and Technology (MNIST) dataset show that the speedups in accumulative training time and each additional training time of the proposed BLS over the existing BLS are 24.81%~ 37.99% and 36.45%~ 58.96%, respectively, and the speedup in total training time is 37.99%. In our experiments, the proposed BLS and the existing BLS both achieve the same testing accuracy when the tiny differences (≤ 0.05%) caused by the numerical errors are neglected, and the above-mentioned tiny differences and numerical errors become zeroes and ignorable, respectively, when the ridge parameter is not too small.

26 May 2025 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
HZHufei Zhu
Hufei Zhu
JYJie You
Jie You
YZYikui Zhai
Yikui Zhai
YXYing Xu
Ying Xu
TWTianlei Wang
Tianlei Wang
CCC.L. Philip Chen
C.L. Philip Chen
With the advancement of industrial automation, surface defect inspection has become crucial for battery manufacturing quality assurance. However, existing public datasets lack sufficient resources for battery surface defects, hindering the process of battery defect inspection. To address this, we developed the Multi-class Battery Surface Defect dataset (M-BSD), which contains 2,491 images and 6,639 samples spanning three defect types. Meanwhile, deep learning-based methods have made progress in industrial quality inspection but face challenges in handling complex battery defect details, defect category similarities, and small target defects. These issues hinder accurate, quantitative analysis of defects. To overcome above challenges, we propose the DS2A-Former, a novel Dual Stream Spatial Attention Transformer Network that integrates semantic and spatial information to enhance segmentation performance under complex conditions. The Dual Scale Spatial Cross Attention (DSSCA) module facilitates information exchange across scales and suppresses background noise, while the Dual Query Gated Spatial Attention (DQGSA) module uses a spatial gating mechanism to prioritize knowledge between two queries, improving representation of defect information. Extensive experiments on the M-BSD dataset validate the effectiveness of our approach. We also experimented it on Crack-500 dataset, demonstrating its broad applicability in defect segmentation across domains. Our code and dataset are available at https://github.com/yikuizhai/DS2A-Former.

25 March 2025 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
YZYikui Zhai
Yikui Zhai
WPWenfeng Pan
Wenfeng Pan
YLYanyang Liang
Yanyang Liang
HZHufei Zhu
Hufei Zhu
ZLZhihao Long
Zhihao Long
FSFabio Scotti
Fabio Scotti
Container marking anomaly detection (CMAD) aims to identify markings that deviate from a standard reference image. Currently, the manual execution of CMAD is inefficient and susceptible to errors. Existing machine vision methods are limited to detecting text-based or specific types of markings, rendering type-agnostic CMAD unattainable. To address this limitation, we propose an innovative framework for CMAD, named Bidirectional Feature Pyramid Siamese Anomaly Detection Network (BiSiNet). BiSiNet consists of the Siamese Encoder (SE), the Bidirectional Feature Pyramid Decoder (BiD), and the Triplet Classifier Prediction Head (TCPH). SE extracts the difference feature pyramid between the target and reference images. BiD decodes the difference feature pyramid in both top-down and bottom-up paths, effectively solving the problem of global information loss during decoding in traditional single-path decoder. TCPH utilizes both local and global features to predict outcomes at both the pixel and patch levels, integrating these results to resolve inconsistencies in predictions. BiSiNet not only possesses strong resistance to disturbances such as reflections and shadows but also has accurate localization capabilities. Due to the limited diversity and the difficulty in obtaining anomalous samples, we designed an innovative Cellular Anomaly Generation Strategy (CAGS). CAGS employs Cellular Noise to generate marking masks, creating realistic pseudo marking images. Through innovative design, it produces four types of samples, thereby enhancing sample diversity and improving training generalization. Extensive experiments using our established ContainerMAD dataset have demonstrated that BiSiNet and CAGS outperform other methods in both container marking anomaly detection and localization. Additionally, favorable outcomes have been achieved on the MvtecAD dataset.