FZFan Zhou

Fan 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

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Fan Zhou has a total of 67 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Mitigating Structural Fragmentation in Sonar Deconvolution: A Directional Total Variation Approach With FISTA Acceleration

    by Pengcheng Pan, Xuesong Liu, Fei Wang, Fan Zhou, Rongxin Jiang, Yaowu Chen

    Underwater 3-D imaging systems face critical challenges in achieving high-resolution reconstructions owing to physical aperture limitations and computational complexity. Conventional beamforming methods suffer from wide mainlobes and elevated sidelobes, whereas existing deconvolution approaches, though effective in super-resolution recovery of discrete point targets, tend to oversegment contiguous scattering bodies into sparse point clusters under operational conditions, leading to structural fragmentation and compromised target detectability. This study proposes a deconvolution framework for underwater targets through a hybrid regularization combining total variation and L1-norm constraints. The comprehensive validation demonstrates superior performance in continuous structural reconstruction, with experimental results showing an improvement in the structural similarity index compared to other deconvolution methods. Field tests conducted in marine environments further validate the framework’s ability to suppress sidelobe artefacts while maintaining contour integrity, thus establishing its potential for high-precision underwater detection and mapping applications.


  • Ranging Enhancement for Underwater Collision Avoidance Sonar via ISAC Waveform and Adaptive Sensing Acceleration

    by Shihui Liang, Weibo Mao, Dongdong Zhao, Peng Chen, Wei Zou, Fan Zhou, Wei Wu, Fuling Huang, Ronghua Liang

    Traditional collision avoidance sonars perform a round-trip ranging measurement per operational cycle, where longer detection distances lead to longer cycle durations. The low propagation speed of acoustic signals increases their susceptibility to Doppler effects, which can ultimately degrade ranging accuracy or success rates and, in severe cases, lead to economic losses or even loss of life. To address this issue, we propose a ranging enhancement method for collision avoidance sonar inspired by integrated sensing and communication (ISAC) principles. At the signal level, we introduce a tail-flipping linear frequency modulated (TF-LFM) waveform along with a Doppler channel estimation method to maintain stability under severe channel conditions. Simulation results demonstrate that the proposed waveform achieves robust measurement under strong Doppler distortions. With a remarkably low average peak-to-average power ratio (PAPR) of approximately 3.01 dB, the signal is easy to transmit and resilient to nonlinear distortion. At the processing level, we propose an adaptive sensing acceleration (ASA) mechanism that allows the sonar to emit new pulses before previous echoes return, thereby reducing ranging latency. In a field pool experiment, the upgraded sonar system achieved a 65.9% reduction in latency per ranging cycle using the proposed method, demonstrating a substantial enhancement in measurement efficiency.


  • Open Access

    Joint Sparse Estimation Method for Array Calibration Based on Fast Iterative Shrinkage-Thresholding Algorithm

    by Boxuan Gu, Xuesong Liu, Fei Wang, Xiang Gao, Fan Zhou

    Existing array calibration methods rely on the geometric characteristics of the array or signal matrix, which limits their flexibility and robustness. This study addresses this challenge by proposing a novel joint sparse estimation method for array gain and phase calibration. By leveraging the sparsity of calibration signals and the dictionary mismatch model, the proposed method, based on the fast iterative shrinkage-thresholding algorithm (FISTA), jointly estimates the discrete on-grid azimuths and continuous off-grid offsets of the direction of arrival (DOA) of calibration signals. The method employs a spatial domain filtering technique based on the maximum a posteriori probability to mitigate the bias induced by phase errors in the calibration signal estimation, enhancing calibration accuracy. Furthermore, the iterative estimation framework was optimized to extend the applicability of the method from linear to uniform planar arrays. The results demonstrated that the root mean squared error (RMSE) of the beam pattern for various array types decreased by one to two orders of magnitude after calibration. Compared with existing state-of-the-art methods, the proposed approach exhibited stable performance and superior estimation accuracy under conventional signal-to-noise ratio conditions. Moreover, the proposed method maintained high stability and lower RMSE as the gain and phase error values increased.


  • Multi-Label Zero-Shot Learning Via Contrastive Label-Based Attention

    by Shixuan Meng, Rongxin Jiang, Xiang Tian, Fan Zhou, Yaowu Chen, Junjie Liu, Chen Shen

    Multi-label zero-shot learning (ML-ZSL) strives to recognize all objects in an image, regardless of whether they are present in the training data. Recent methods incorporate an attention mechanism to locate labels in the image and generate class-specific semantic information. However, the attention mechanism built on visual features treats label embeddings equally in the prediction score, leading to severe semantic ambiguity. This study focuses on efficiently utilizing semantic information in the attention mechanism. We propose a contrastive label-based attention method (CLA) to associate each label with the most relevant image regions. Specifically, our label-based attention, guided by the latent label embedding, captures discriminative image details. To distinguish region-wise correlations, we implement a region-level contrastive loss. In addition, we utilize a global feature alignment module to identify labels with general information. Extensive experiments on two benchmarks, NUS-WIDE and Open Images, demonstrate that our CLA outperforms the state-of-the-art methods. Especially under the ZSL setting, our method achieves 2.0% improvements in mean Average Precision (mAP) for NUS-WIDE and 4.0% for Open Images compared with recent methods.


  • ESOD: Efficient Small Object Detection on High-Resolution Images

    by Kai Liu, Zhihang Fu, Sheng Jin, Ze Chen, Fan Zhou, Rongxin Jiang, Yaowu Chen, Jieping Ye

    Enlarging input images is a straightforward and effective approach to promote small object detection. However, simple image enlargement is significantly expensive on both computations and GPU memory. In fact, small objects are usually sparsely distributed and locally clustered. Therefore, massive feature extraction computations are wasted on the non-target background area of images. Recent works have tried to pick out target-containing regions using an extra network and perform conventional object detection, but the newly introduced computation limits their final performance. In this paper, we propose to reuse the detector’s backbone to conduct feature-level object-seeking and patch-slicing, which can avoid redundant feature extraction and reduce the computation cost. Incorporating with a sparse detection head, we are able to detect small objects on high-resolution inputs (e.g., 1080P or larger) for superior performance. The resulting Efficient Small Object Detection (ESOD) approach is a generic framework, which can be applied to both CNN- and ViT-based detectors to save the computation and GPU memory costs. Extensive experiments demonstrate the efficacy and efficiency of our method. In particular, our method consistently surpasses the SOTA detectors by a large margin (e.g., 8% gains on AP) on the representative VisDrone, UAVDT, and TinyPerson datasets. Code will be made public soon.


  • Expansion of Field of View for Near-Field Fast Beamforming in 3-D Acoustic Imaging Based on the Optimized Subregion Approach

    by Fei Wang, Xuesong Liu, Chenyi Lin, Xiang Gao, Boxuan Gu, Fan Zhou, Yaowu Chen

    Real-time underwater 3-D acoustic imaging employs various fast-beamforming methods that significantly reduce the computational cost. However, in the near-field region, these methods rely on a popular model based on the Fresnel approximation, which has a narrow field of view (FOV) boundary of approximately $26^\circ$ . The FOV of the near-field region is very limited compared with that of the far-field region. Therefore, in this study, an optimized subregion approach is proposed to eliminate the FOV limitation for fast beamforming to improve the FOV of the near-field region. First, the FOV is divided into subregions, and within each subregion, a linear approximation is adopted to simplify the time-delay expression, with the approximation error limited to a reasonable threshold. Furthermore, the least-squares method and coordinate rotation techniques are employed, and the FOV for each subregion is reshaped to an ideal shape. Subsequently, a nested nonuniform fast Fourier transform is proposed to implement fast beamforming, and subregions can be computed in parallel. The results demonstrate that the proposed approach overcomes the limited FOV that exists for near-field fast beamforming in 3-D acoustic imaging and has a computational complexity comparable with those of existing algorithms. In addition, this approach supports an irregular planar array and maintains a satisfactory performance.

  • Contrastive visual feature filtering for generalized zero-shot learning

    by Shixuan Meng, Rongxin Jiang, Xiang Tian, Fan Zhou, Yaowu Chen, Junjie Liu, Chen Shen

  • Open Access

    A Unified Asymmetric Knowledge Distillation Framework for Image Classification

    by Xin Ye, Xiang Tian, Bolun Zheng, Fan Zhou, Yaowu Chen

    Knowledge distillation is a model compression technique that transfers knowledge learned by teacher networks to student networks. Existing knowledge distillation methods greatly expand the forms of knowledge, but also make the distillation models complex and symmetric. However, few studies have explored the commonalities among these methods. In this study, we propose a concise distillation framework to unify these methods and a method to construct asymmetric knowledge distillation under the framework. Asymmetric distillation aims to enable differentiated knowledge transfers for different distillation objects. We designed a multi-stage shallow-wide branch bifurcation method to distill different knowledge representations and a grouping ensemble strategy to supervise the network to teach and learn selectively. Consequently, we conducted experiments using image classification benchmarks to verify the proposed method. Experimental results show that our implementation can achieve considerable improvements over existing methods, demonstrating the effectiveness of the method and the potential of the framework.


  • Deforming Garment Classification With Shallow Temporal Extraction and Tree-Based Fusion

    by Li Huang, Tong Yang, Rongxin Jiang, Xiang Tian, Fan Zhou, Yaowu Chen

    A novel RGB-based continuous perception garment classification approach is proposed in this letter, with the aim of identifying the correct category of the garment from a set of categories. It has been observed that treating a video of the continuous deformation of cloth as a set of disordered static figures leads to low classification precision. On the contrary, investigating the temporal information between frames can significantly improve the quality of extracted features and increase classification performance. In this regard, we propose a hybrid temporal fusion RGB-based algorithm, including an improved image-level shallow temporal feature extraction module (STEM) and a binary-tree fusion module (BiTF) for adaptive feature fusion. STEM incorporates multi-scale optical flow and long-short-term memorised information to capture both static features in every single image and dynamic features in consecutive images. BiTF constructs a tree-shaped structure to fuse an arbitrary number of extracted features in a video. The proposed algorithm achieves 97.5% accuracy on the public dataset, surpassing the state-of-the-art RGB-based classification algorithm by 44.0% and achieving comparable performance with RGBD-based algorithms.