HCHongnan Cheng

Hongnan Cheng

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Hongnan Cheng has a total of 18 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Learning invariant causal logic for traffic flow forecasting: A disentangled spatiotemporal intervention approach

    by Chaozhi Yu, Hongnan Cheng, Zhizhe Lin, Heng Zhang, Teng Zhou, Xin Zhao, Chunjie Cao

  • EAL-ICNet: Enhanced Attention Lightweight Network With Multi-Scale Transformer Modules for IC Image Segmentation

    by Hongnan Cheng, Chaozhi Yu, Yan Zhang, Ziwen Yuan, Chenguang Zhang

    The rapid development of deep learning, especially convolutional neural networks, has greatly enhanced automated segmentation of integrated circuit (IC) images. However, current methods face challenges: 1) Large-scale industrial IC images require high-throughput models, which many high-parameter networks struggle to meet. 2) Conventional convolutions are limited in capturing global dependencies within circuit structures. 3) Optical microscopy images often contain noise and artifacts that hinder accurate segmentation. To address these issues, we propose EAL-ICNet (Enhanced Attention Lightweight IC Network), a lightweight architecture designed for IC image segmentation. EAL-ICNet integrates a Depthwise Asymmetric Bottleneck (DAB) module, combining depthwise asymmetric and dilated convolutions for improved efficiency and reduced complexity. Inspired by Transformer mechanisms, Token Statistics Self-Attention (TSSA) and Parallel Multi-Scale Attention (PMA) are introduced to capture non-local dependencies and suppress noise. EAL-ICNet also employs multi-scale contextual modeling for better feature extraction. We evaluate the model on four real IC datasets and demonstrate its ability to balance accuracy and inference speed, making it suitable for industrial applications. The code is available at https://github.com/chndaz/EAL-ICNet.

  • Quantum-Inspired Dynamic Spatiotemporal Matching Transformer for Traffic Flow Forecasting

    by Chaozhi Yu, Zhizhe Lin, Hongnan Cheng, Chunjie Cao, Teng Zhou, Man-Fai Leung

    Accurate spatiotemporal traffic modeling and prediction are essential for intelligent transportation systems (ITS). Conventional models struggle with complex spatiotemporal variation and dynamic traffic patterns. The spatiotemporal heterogeneity of different road networks is challenging, which limits the models to accurately capture traffic flow patterns. In this paper, we propose a quantum-inspired dynamic spatiotemporal matching transformer (DSTMT) to achieve more robust and accurate traffic flow forecasting. We first design a temporal pattern embedding to capture temporal heterogeneity to improve the accuracy of traffic flow forecasting. Then, we design a memory-augmented pattern matching to extract the representative traffic pattern for each traffic node. We also develop a quantum-inspired adaptive transfer mechanism to mitigate the error caused by the discrepancy between historical traffic flow and forecasting. After that, we introduce an iTransformer to capture the spatial heterogeneity for each traffic node to improve prediction performance. Extensive experiments are conducted on five real-world datasets compared with 21 state-of-the-art methods and 7 baselines. The experimental results demonstrate the superior performance of DSTMT for traffic flow forecasting.


  • Multi-Scale Cross-Dimensional Attention Network for Gland Segmentation

    by Chaozhi Yu, Hongnan Cheng, Yufei Huang, Zhizhe Lin, Teng Zhou

    Gland lesions affect a large global population. Accurately segmenting surface structures is crucial for assisting in the diagnosis of these diseases. In this direction, we investigate two key issues: 1) How to accurately segment gland morphology and irregular boundaries and 2) How to distinguish gland internal heterogeneity and its similarity to the background. The main results are that 1) parallel multi-scale attention (PMA) smooths the segmentation of blurred boundaries of varying sizes and improves detail accuracy. 2) Cross-dimensional attention (CDA) models the dependencies between gland channels and spatial dimensions to enhance the understanding of spatial information both inside and outside the gland, thereby more accurately distinguishing the gland from the background. Per the main results, we propose a multi-scale cross-dimensional attention network (MCANet) for gland segmentation. Extensive experiments on six real-world datasets demonstrate the superior performance of our method in gland segmentation. The source code is available at https://github.com/yuchaozhi/MCANet.


  • Open Access

    Iterative deep PU learning combined with data features

    by Hongnan Cheng, Yan Zhang, Chaozhi Yu, Boyu Li, Chenguang Zhang

    In many real-world applications, learning a binary classifier relying solely on positive and unlabeled (PU) training sample data is an important and challenging task. The absence of negative training samples leads to the necessity for PU learning methods to extract effective additional information from unlabeled training samples. This paper proposes a method that combines Mean Teacher and Smooth Neighbors on Teacher Graphs (SNTG) to address the issue of model overfitting towards unlabeled data in PU learning, where unlabeled data is typically assigned a lower weight. The use of the Mean Teacher model ensures effective utilization of unlabeled data information. SNTG maps the data distribution from high-dimensional space to low-dimensional manifolds or clustering structures, ensuring that the model captures the complex local structures between data points. Moreover, when using the Mean Teacher for data augmentation of unlabeled data points, the intrinsic relationships between the data points are often overlooked, and SNTG helps to address this limitation. By combining the SNTG loss, consistency loss in the Mean Teacher, and non-negative PU loss, the effectiveness of the proposed framework is validated through experiments on three benchmark datasets (MNIST, Fashion-MNIST, CIFAR-10) and two real-world datasets (Avila, EGSS—Electrical Grid Stability Simulated). The results show accuracies of 95.03%, 95.18%, 89.73%, 81.12%, and 92.55%, surpassing the performance of most state-of-the-art PU learning algorithms.


  • Segmentation of IC Images in Integrated Circuit Reverse Engineering Using EfficientNet Encoder Based on U‐Net++ Architecture

    by Hongnan Cheng, Chaozhi Yu, Chenguang Zhang

    Segmentation of electrical components and metal traces from integrated circuit (IC) images is crucial for IC reverse engineering. Existing image segmentation methods face significant challenges when applied to IC images, including high resolution, limited training data, and the need for precise segmentation. To address these issues, this study proposes a combined approach of segmentation and post‐processing. During the segmentation stage, we use UNet++ as the base architecture, with EfficientNet‐B7 as the encoder, resulting in an E‐UNet++ model. This model effectively combines the efficiency and pre‐training capabilities of EfficientNet with the ability of UNet++ to capture both global structural information and fine‐grained boundary details in IC images, enabling it to effectively handle challenges such as high resolution and limited training samples. In the post‐processing stage, to address potential noise caused by the insufficient utilization of spatial location information in network‐based methods, we propose the use of Hough circle detection and median filtering to eliminate noise from vias and non‐via regions. Compared to the suboptimal segmentation model, our proposed method achieved a 0.58% improvement in mean intersection over union (mIoU) and a 0.33% improvement in mean pixel accuracy (MPA) on the real‐world dataset and a 0.78% improvement in mIoU and a 0.44% improvement in MPA on the open‐source dataset. These experimental results demonstrate that our method effectively improves the accuracy of IC segmentation.


  • Open Access

    Glandular Tissue Segmentation Based on EMA-Swin UNet Model

    by Hongnan Cheng, Chaozhi Yu, Yan Zhang, Boyu Li, Wenxiang Huang, Chenguang Zhang

    The accurate gland segmentation from digitized H&E (hematoxylin and eosin) histology images with a wide range of histologic grades of cancer is quite challenging. In recent years, several fully convolutional network methods have been proposed, with UNet being the most classic. The UNet model, with its symmetric structure, has shown excellent performance in gland segmentation tasks. However, the locality of convolution operations in UNet also limits its ability to capture global dependencies. To address this limitation, this paper proposes a novel deep glandular tissue image segmentation network based on Swin UNet, termed EMA-Swin UNet. This network replaces CNN modules with Swin Transformer modules to capture both local and global representations. Additionally, the EMA-Swin UNet incorporates an Efficient Multi-scale Attention (EMA) module to enhance multi-scale feature extraction for glandular tissues of various sizes by capturing global dynamic features and long-range smooth features from the encoder outputs. By integrating edge-detection pooling, we enhanced the refinement of prediction maps produced by the EMA-Swin UNet. Moreover, we standardized the staining across both the ClaS dataset and the six-grade tumor differentiation dataset from EBHI-Seg using Reinhard normalization. The final segmentation results are compared with those of classical gland segmentation algorithms on the ClaS and EBHI-Seg datasets, demonstrating the effectiveness of our proposed method. Particularly, on the GlaS dataset, the mDice reached 0.894.


  • Open Access

    Segmentation of IC images in integrated circuit reverse engineering using EfficientNet encoder based on U-Net++ architecture

    by Hongnan Cheng, Chaozhi Yu, Chenguang Zhang

    The chip industry is essential for national security and economic development, with integrated circuit (IC) reverse engineering playing a vital role in analyzing chip structures. This process involves several steps, including layer-by-layer image acquisition using scanning electron microscopy (SEM), device identification, gate net extraction, and function inference. Segmenting electrical components and metal lines from IC images is crucial for these analyses. However, traditional image segmentation methods often fail to handle the complex and variable conditions of IC images due to insufficient expert knowledge. This study introduces an improved approach, using the UNet ++ architecture and effentnet-b7 as the encoder, called the E- UNet++ model. A post-processing denoising stage is added that contains Hough circle detection and median filtering for extracting metal lines and perforations in IC images. The primary contributions of this method are: (1) it enables fully automatic detection of metal lines and vias without manual intervention, and (2) it combines E-UNet++, Hough circle detection, and median filtering in a hybrid approach to accurately locate metal lines and vias. Experimental results on over ten thousand IC images, each measuring 1024×1024 and provided by a company, show that training with just 393 images allows the E-UNet++ model to effectively segment metal lines and vias. The average intersection over union (mIoU) is 98.09% and the mean pixel accuracy (MPA) is 99.06%, surpassing the performance of existing methods.