WZWenzhe Zhai

Wenzhe Zhai

*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


Wenzhe Zhai has a total of 35 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Top Cited Papers

    Privacy-Preserving Object Counting in Lightweight Intelligent Transportation System

    by Wenzhe Zhai, Xianglei Xing, Jing-An Cheng, Gwanggil Jeon

    In Intelligent Transportation Systems (ITS), tiny machine learning enables resource-efficient on-device learning and deployment of AI models for critical tasks such as traffic flow prediction and pedestrian safety. However, existing object counting models typically consist of a large number of parameters, which require high computational costs that are unsuited for on-device learning. Moreover, these models are usually trained in a centralized manner, which raises data privacy concerns and relies on the unrealistic assumption of independent and identically distributed (IID) data across all vehicles and environments. To address these challenges, we propose a lightweight network, termed Privacy-aware Efficient Generalization Network (PEGNet), that preserves user privacy and improves cross-scenario generalization. Specifically, this network incorporates a Lightweight Adaptive Enhancement (LAE) module with deformable convolutions to efficiently capture multi-scale features under strict resource constraints. We further employ a federated learning framework to collaboratively train the model across distributed IoV nodes with limited local data, thereby significantly reducing per-node computation while safeguarding user data privacy. To enhance the robustness of non-IID data, we introduce a Moment Distance loss (MDloss) that prompts the model to adapt to diverse traffic scenes. Experiments on multiple vehicle and crowd counting benchmarks demonstrate that the proposed method achieves superior efficiency and accuracy.


  • Composed image retrieval by Multimodal Mixture-of-Expert Synergy

    by Wenzhe Zhai, Mingliang Gao, Gwanggil Jeon, Qiang Zhou, David Camacho


  • Knowledge Generation and Distillation for Road Segmentation in Intelligent Transportation Systems

    by Jianyong Wang, Mingliang Gao, Wenzhe Zhai, Imad Rida, Xianxun Zhu, Qilei Li

    The rapid development of generative AI has opened a new era in Intelligent Transportation Systems (ITS). However, deploying high-performance segmentation models on resource-constrained edge devices remains challenging due to their substantial computational demands. To address this problem, in this work, we propose a lightweight road segmentation framework termed Knowledge Generation and Distillation (KGD). In the KGD, a lightweight Student model learns from a high-precision Teacher model. This approach balances accuracy and computational efficiency. To further enhance the knowledge transfer process, we introduce a knowledge distillation loss to better supervise the discrepancy between the Teacher and Student models. Meanwhile, we incorporate graph convolution to capture complex spatial dependencies. This can effectively enhance the understanding of road structure and irregular boundaries. Additionally, we built a Multi-scale Lightweight Spatial Attention (MS-LSA) module to focus on multi-scale spatial road information. Experimental results demonstrate that the proposed KGD achieves 96.33% and 94.02% in Max F1-measure and average precision(AP) on KITTI-Road, with only 1.17M in parameters and scores 6.73ms/frame in inference speed. It achieves a superior balance between accuracy and efficiency compared to mainstream deep networks. These advantages make KGD suitable for real-time use in large-scale ITS applications, such as smart traffic monitoring, autonomous vehicle perception, and adaptive traffic control systems.


  • Cross-Camera Discriminative Person Association by Unsupervised Frame Clustering and Selection

    by Qilei Li, Mingliang Gao, Guisheng Zhang, Wenzhe Zhai, Gwanggil Jeon

    The objective of cross-camera persson association is to identify individuals captured across disjoint cameras. It is achieved by Re-Identification (ReID) models, which extract unique identity representations from the visual input, dominated by video sequence. Most ReID methods primarily focus on modifying the backbone network architecture to learn more representative features of people. However, these methods often overlook the impact of low-quality frames on the training process. Several studies have confirmed that low-quality data not only hinders the model from learning meaningful content but also diminishes its performance. One possible solution is to manually label the quality of each frames, but this is time-consuming and inefficient. In this paper, we propose a Unsupervised Frame Clustering and Selection framework called UFCS to address this problem by applying the unsupervised clustering to automatically select high-quality frames. Specifically, we applied three unsupervised clustering solutions for high-quality frame selection, namely K-means, Deep K-means, and DBSCAN. These clustering techniques integrate both appearance and, indirectly, temporal consistency by operating within tracklets. These schemes perform clustering in the image or deep feature space to select highquality frames for network training. This straightforward yet effective approach enables the ReID network to generate a more discriminative representations, thereby improving recognition performance. Experimental results obtained on the challenging video-based person ReID datasets MARS indicate that our proposed scheme can outperform related state-of-the-art methods by a large margin.


  • Open Access

    Shackled Dancing: A Bit-Locked Diffusion Algorithm for Lossless and Controllable Image Steganography

    by Tianshuo Zhang, Gao Jia, Wenzhe Zhai, Rui Yann, Xianglei Xing

    Data steganography aims to conceal information within visual content, yet existing spatial- and frequency-domain approaches suffer from trade-offs between security, capacity, and perceptual quality. Recent advances in generative models, particularly diffusion models, offer new avenues for adaptive image synthesis, but integrating precise information embedding into the generative process remains challenging. We introduce Shackled Dancing Diffusion, or SD$^2$, a plug-and-play generative steganography method that combines bit-position locking with diffusion sampling injection to enable controllable information embedding within the generative trajectory. SD$^2$ leverages the expressive power of diffusion models to synthesize diverse carrier images while maintaining full message recovery with $100\%$ accuracy. Our method achieves a favorable balance between randomness and constraint, enhancing robustness against steganalysis without compromising image fidelity. Extensive experiments show that SD$^2$ substantially outperforms prior methods in security, embedding capacity, and stability. This algorithm offers new insights into controllable generation and opens promising directions for secure visual communication.

  • Training-Free 3-D Face Avatars Generation by Knowledge Discovering in Foundational Models

    by Qilei Li, Mingze Xu, Wenzhe Zhai, David Camacho, Gwanggil Jeon

    An informative 3-D avatar, closely mirroring real-world traits, plays a pivotal role in accessing the metaverse. Traditional methods for creating 3-D avatars usually employ one-to-one training, which restricts avatar diversity. To enhance style diversity in generated 3-D avatars, we utilize synthesized images derived with prompts from ChatGPT in a conversational manner, ultimately resulting in a broader range of 3-D variations. Rather than creating models from scratch, we devise a training-free framework that utilizes established large-scale foundation models. Specifically, we employ a real-world image synthesis technique guided by text prompts that are generated by ChatGPT in a conversational manner, to describe the desired characteristics of the synthesized image. As a result, these informative latent representations can accurately reflect the distinct style of the synthesized image, and further lead to the creation of photorealistic and diverse 3D avatars. Our training-free design allows this proposed method to achieve competitive performance compared to existing generation models, while requiring minimal computational resources.


  • Zero-Shot Object Counting With Vision-Language Prior Guidance Network

    by Wenzhe Zhai, Xianglei Xing, Mingliang Gao, Qilei Li

    The majority of existing counting models are designed to operate on a singular object category, such as crowds or vehicles. The emergence of multi-modal foundational models, e.g ., Contrastive Language-Image Pre-training (CLIP), has paved the way for class-agnostic counting. This approach facilitates the counting of objects across diverse classes within a single image based on textual indications. However, class-agnostic counting models based on CLIP confront two primary challenges. Firstly, the CLIP model exhibits limited sensitivity towards location information, which prioritizes global content over the precise localization of objects. Therefore, directly employing the CLIP model is regarded as suboptimal. Secondly, these models commonly employ frozen pre-trained vision and language encoders while disregarding potential misalignment within the constructed hypothesis space. In this paper, we propose a unified framework, named the Vision-Language Prior Guidance (VLPG) Network, to tackle these two challenges. The VLPG consists of three key components, namely the Grounding DINO module, Spatial Prior Calibration (SPC) module, and Object-Centric Alignment (OCA) module. The Grounding DINO module utilizes the spatial-awareness capability of extensive pre-trained object grounding models to incorporate the spatial position as an additional prior for a particular query class. This adaptation enables the network to concentrate more precisely on the exact location of the objects. Meanwhile, the SPC module is built to extract the long-range dependencies and local regions of the spatial position. Additionally, to align the feature space across different modalities, we design an OCA module that condenses textual information into an object query which serves as an instruction for cross-modality matching. Through the collaborative efforts of these three modules, multimodal representations are aligned while maintaining their discriminative nature. Comprehensive experiments conducted on various benchmarks validate the effectiveness of the proposed model.


  • Multi-View Gait Recognition With Joint Local Multi-Scale and Global Contextual Spatio-Temporal Features

    by Wenzhe Zhai, Haomiao Li, Chaoqun Zheng, Xianglei Xing

    Existing gait recognition methods are capable of extracting rich spatial gait information but often overlook fine-grained temporal features within local regions and temporal contextual information across different sub-regions. Considering gait recognition as a fine-grained recognition task and each individual exhibits uniqueness in their movements across different temporal sequences, we propose a local multi-scale and global contextual spatio-temporal (LMGCS) network for gait recognition. It divides the whole gait sequence into sub-sequences with multiple spatio resolutions and extracts multi-scale temporal features. We extract the temporal context information of different sub-sequences with the transformer, and all sub-sequences are fused to form global features. Furthermore, the loss function that combines the triplet loss function and cross-entropy loss function is utilized to prompt the proposed model to fulfill the gait recognition. The proposed method achieved state-of-the-art results on two popular public datasets. It achieved rank-1 accuracy of 98.0%, 95.4%, and 85.0% on the three walk states of the CASIA-B dataset and 90.9% on the OU-MVLP dataset.


  • Open Access

    CT and MRI Image Fusion via Coupled Feature-Learning GAN

    by Qingyu Mao, Wenzhe Zhai, Xiang Lei, Zenghui Wang, Yongsheng Liang

    The fusion of multimodal medical images, particularly CT and MRI, is driven by the need to enhance the diagnostic process by providing clinicians with a single, comprehensive image that encapsulates all necessary details. Existing fusion methods often exhibit a bias towards features from one of the source images, making it challenging to simultaneously preserve both structural information and textural details. Designing an effective fusion method that can preserve more discriminative information is therefore crucial. In this work, we propose a Coupled Feature-Learning GAN (CFGAN) to fuse the multimodal medical images into a single informative image. The proposed method establishes an adversarial game between the discriminators and a couple of generators. First, the coupled generators are trained to generate two real-like fused images, which are then used to deceive the two coupled discriminators. Subsequently, the two discriminators are devised to minimize the structural distance to ensure the abundant information in the original source images is well-maintained in the fused image. We further empower the generators to be robust under various scales by constructing a discriminative feature extraction (DFE) block with different dilation rates. Moreover, we introduce a cross-dimension interaction attention (CIA) block to refine the feature representations. The qualitative and quantitative experiments on common benchmarks demonstrate the competitive performance of the CFGAN compared to other state-of-the-art methods.