WWWeibo Wei

Weibo Wei

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Weibo Wei has a total of 76 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • PottsSAGNet: A Potts Model-Based Segmentation Network with Self-Attention and Gaussian Regularization

    by He Xiong, Guangyu Yang, Junli Zhao, Weibo Wei, Zhenkuan Pan, Xuecheng Tai

  • Open Access

    DUM-Net: A model-driven unrolled framework for multi-modal image registration

    by Moyan Cao, Xueying Li, Guangyu Yang, Junli Zhao, Weibo Wei, Jinming Duan, Zhenkuan Pan

    Multi-modal image registration (MIR) remains a challenging problem due to severe appearance discrepancies across imaging modalities and the limited adaptability of handcrafted similarity measures. While deep learning methods have shown strong empirical performance, many existing architectures do not explicitly incorporate physical deformation priors. In this work, we propose DUM-Net, a model-driven deep unrolling framework for multi-modal image registration. Starting from a variational MIR model with a similarity term and a regularization term, we unroll a time-discrete gradient descent process into a layer-wise trainable architecture. DUM-Net consists of three components: an initialization module (IM), an adaptive similarity module (ASM), and a linear elasticity module (LEM). The IM provides deformation initialization to accelerate convergence. The ASM employs a deep neural network to learn an adaptive descent direction for cross-modal matching, thereby overcoming the limitations of handcrafted similarity functionals. The LEM introduces a parameterized linear-elastic prior during forward propagation to encourage smooth and physically plausible deformation fields. In this way, DUM-Net combines the flexibility of learned similarity modeling with the structural guidance of variational unrolling and explicit physical regularization. Experiments on T1/T2 MR, CT/MR, and FL/IR registration tasks demonstrate that DUM-Net consistently outperforms existing methods, achieving Dice score improvements of 3.74% for small and 6.16% for large deformations on T1/T2 MR image pairs, together with strong performance on CT/MR and FL/IR datasets. The source code is publicly available at https://github.com/QDUJackson/DUM-Net.

  • Sperf: A trainable smooth non-monotonic activation function with enhanced nonlinearity

    by Xueying Li, Moyan Cao, Junli Zhao, Weibo Wei, Zhenkuan Pan

    ReLU, as the core nonlinear unit in numerous deep learning backbones, plays a crucial role in this field. In recent years, various smooth ReLU-Like activation functions have demonstrated superior performance across different tasks and network architectures. In this paper, we propose a novel trainable ReLU-Like activation function, Sperf, constructed by multiplying Softplus with the Gaussian error function (erf). Sperf enhances nonlinear representation capability while maintaining numerical stability. Experimental results show that networks equipped with Sperf achieve significant performance gains across diverse tasks: six image-classification datasets; Cityscapes for semantic segmentation; and MICCAI BraTS2020 for unsupervised image registration.


  • ReLU-like non-monotonic smooth activation functions based on regularized heaviside functions and extensions

    by Pengfei Ren, Tony Yuxiang Pan, Guangyu Yang, Yanchen Guo, Weibo Wei, Zhenkuan Pan

    ReLU is a widely adopted activation function in neural networks of deep learning for different application tasks. In order to overcome its limitations of non-smoothness and dying neuron problems, various improved versions with non-monotonic, smooth properties have been proposed via different innovative methods. In this work, we present a systematic method to construct this kind of activation functions via regularized S-shaped Heaviside functions based on exponential function, logarithmic function, trigonometric function, and Gaussian error function, etc.. The derived ReLU-like activation functions have the common properties of nonlinearity, infinite differentiability, non-monotonicity, zero-centering. Based the requirements of S-shaped regularized Heaviside functions, we modify some existing classical Logish, Smish, and CaLU to ReLU-like activation functions to approach ReLU at positive or negative regions to inherit the merits of ReLU. Additionally, we extend the proposed method to activation functions composed of two and three piecewise linear functions, and present some examples, such as Leaky ReLU, V-shaped ReLU, projection formula, soft and hard thresholding formulas. This method can be utilized to transform the activation functions with multiple piecewise linear functions into regularized ones naturally.


  • A trainable variational Chan-Vese network based on algorithm unfolding for image segmentation

    by Zhaolong Cui, Tony Yuxiang Pan, Guangyu Yang, Junli Zhao, Weibo Wei

    Deep neural networks based on algorithm unfolding have received significant attention in computer vision in recent years due to their advantages in explainability and efficiency compared with traditional model-based methods and data-driven deep neural networks. In this paper, we propose a novel variational neural network architecture for image segmentation, inspired by the well-known Chan-Vese (CV) model, and the recent PottSMGNet and Double-Well Net proposed in [ 18 , 31 ]. First, we reformulate the CV model into a vectorial version (VCV) by applying multi-channel convolution to the original images for feature recognition. Then, we incorporate multi-channel Fields of Experts, employing total variation for contour length regularization. The former is designed based on trainable low-pass filters, while the latter is constructed using a convex combination of discrete cosine transform bases. Next, we decompose the VCV model into sub-problems of optimization using alternating-direction iterative minimization methods. This involves introducing auxiliary variables and learnable penalty parameters to address each sub-problem effectively. Finally, we design an end-to-end deep neural network called CVN (CV model-based neural network) by establishing correspondences between iterative algorithms and deep neural networks. The activation functions of CVN employ parametric soft thresholding formulas and projection formulas expressed as classical rectified linear unit (ReLU) functions. We conduct experiments on various ground truth datasets to demonstrate the visual properties of CVN and highlight its advantages over the original CV model and other classical deep learning models.


  • Open Access

    A Variational Neural Network Based on Algorithm Unfolding for Image Blind Deblurring

    by Shaoqing Gong, Yeran Wang, Guangyu Yang, Weibo Wei, Junli Zhao, Zhenkuan Pan

    Image blind deblurring is an ill-posed inverse problem in image processing. While deep learning approaches have demonstrated effectiveness, they often lack interpretability and require extensive data. To address these limitations, we propose a novel variational neural network based on algorithm unfolding. The model is solved using the half quadratic splitting (HQS) method and proximal gradient descent. For blur kernel estimation, we introduce an L0 regularizer to constrain the gradient information and use the fast fourier transform (FFT) to solve the iterative results, thereby improving accuracy. Image restoration is initiated with Gabor filters for the convolution kernel, and the activation function is approximated using a Gaussian radial basis function (RBF). Additionally, two attention mechanisms improve feature selection. The experimental results on various datasets demonstrate that our model outperforms state-of-the-art algorithm unfolding networks and other blind deblurring models. Our approach enhances interpretability and generalization while utilizing fewer data and parameters.


  • PottsNN: A Variational Neural Network Based on Potts Model for Image Segmentation

    by Yeran Wang, ZhengHong Zhong, Junli Zhao, Shaoqing Gong, Zhenkuan Pan, Weibo Wei

  • HyperSINet: A Synergetic Interaction Network Combined With Convolution and Transformer for Hyperspectral Image Classification

    by Qixing Yu, Weibo Wei, Dantong Li, Zhenkuan Pan, Chenyu Li, Danfeng Hong

    In hyperspectral images (HSIs), both local and non-local features play crucial roles in classification tasks. Vision Transformer (VIT) can extract non-local features through attention mechanisms, while Convolutional Neural Networks (CNN) excel at handling local components. However, in traditional dual-branch models based on VIT and CNN, there is a lack of interaction during feature processing, leading to potential compatibility issues when merging the two types of features. In this article, we propose HyperSINet, a Synergetic Interaction Network that combines VIT and CNN to establish interaction between the two branches, enabling mutual compensation between local and non-local features during the training process and ultimately enhancing the performance of classification tasks. Specifically, we devise a pair of interactors, namely Conv2Trans and Trans2Conv, which serve as intermediaries between the two branches, enabling the VIT branch to refine its local details, while allowing the CNN branch to process larger receptive field non-local features. Typical feature Maps are implemented to visualize the function of the interactors. Furthermore, within the VIT branch, a VIT Encoder with the local mask is developed to strike a balance between emphasizing non-local features and preserving local details, while a lightweight CNN block is designed to process spectral and spatial features in the CNN branch. Extensive experiments conducted on four real-world datasets demonstrate that, under a reasonable count of parameters, HyperSINet surpasses several current state-of-the-art methods.


  • Open Access

    Unsupervised image denoising based on self-attention mechanism

    by Qingmeng Guo, Weibo Wei, Nianrun Xi, Bo Wang

  • GPF-Net: Graph-Polarized Fusion Network for Hyperspectral Image Classification

    by Qixing Yu, Weibo Wei, Zhenkuan Pan, Jingfei He, Shaohua Wang, Danfeng Hong

    Recently, there has been growing interest in hyperspectral images (HSIs) classification tasks, with both Graph Neural Networks (GNN) and Convolutional Neural Networks (CNN) proving to be effective means of analysis. GNN can better capture the spatial structure of HSIs in large target irregular regions through superpixel segmentation, while CNN can refine classification tasks by processing pixel-level features in small target regular regions. However, neither GNN nor CNN models alone can simultaneously consider superpixel-level and pixel-level features to cover both large and small target regions. To fully utilize the strengths of GNN and CNN, we propose a novel model called the Graph-Polarized Fusion Network (GPF). The GPF consists of two branches: the Fusion Graph Neural Network (FGNN) classifier in the GNN branch conducts feature learning on large, irregular target regions using both Graph Convolutional Network (GCN) and Graph Attention Network (GAT) as feature extraction operators. The features are integrated using three aggregators, namely Min, Max, and Weighted Add, followed by updating the nodes through 2D convolutional layers. The Polarized Neural Network (PNN) classifier of the CNN branch primarily works on small, target regular regions using Polarized Self-Attention (PSA) to conduct high-resolution processing on the two dimensions of space and channel without increasing time loss. Additionally, GPF employs residual connections to extract features from long distances and multi-angles. It also uses weighted fusion to integrate the superpixel-level and pixel-level features obtained from the two branches. Rigorous experiments on five real datasets demonstrate that GPF can fully mine the latent features of HSIs, achieving competitive results compared with other state-of-the-art methods.