BSBinghua Su

Binghua Su

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Binghua Su has a total of 31 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    SE-MTCAELoc: SE-Aided Multi-Task Convolutional Autoencoder for Indoor Localization with Wi-Fi

    by Yongfeng Li, Juan Huang, Yuan Yao, Binghua Su

    Indoor localization finds wide-ranging applications in user navigation and intelligent building systems. Nevertheless, signal interference within complex indoor environments and challenges regarding localization generalization in multi-building and multi-floor scenarios have restricted the performance of traditional localization methods based on Wi-Fi fingerprinting. To tackle these issues, this paper presents the SE-MTCAELoc model, a multi-task convolutional autoencoder approach that integrates a squeeze-excitation (SE) attention mechanism for indoor positioning. Firstly, the method preprocesses Wi-Fi Received Signal Strength (RSSI) data. In the UJIIndoorLoc dataset, the 520-dimensional RSSI features are extended to 576 dimensions and reshaped into a 24 × 24 matrix. Meanwhile, Gaussian noise is introduced to enhance the robustness of the data. Subsequently, an integrated SE module combined with a convolutional autoencoder (CAE) is constructed. This module aggregates channel spatial information through squeezing operations and learns channel weights via excitation operations. It dynamically enhances key positioning features and suppresses noise. Finally, a multi-task learning architecture based on the SE-CAE encoder is established to jointly optimize building classification, floor classification, and coordinate regression tasks. Priority balancing is achieved using weighted losses (0.1 for building classification, 0.2 for floor classification, and 0.7 for coordinate regression). Experimental results on the UJIIndoorLoc dataset indicate that the accuracy of building classification reaches 99.57%, the accuracy of floor classification is 98.57%, and the mean absolute error (MAE) for coordinate regression is 5.23 m. Furthermore, the model demonstrates exceptional time efficiency. The cumulative training duration (including SE-CAE pre-training) is merely 9.83 min, with single-sample inference taking only 0.347 milliseconds, fully meeting the requirements of real-time indoor localization applications. On the TUT2018 dataset, the floor classification accuracy attains 98.13%, with an MAE of 6.16 m. These results suggest that the SE-MTCAELoc model can effectively enhance the localization accuracy and generalization ability in complex indoor scenarios and meet the localization requirements of multiple scenarios.


  • Open Access

    High-Resolution Time-Frequency Feature Enhancement of Bowhead Whale Calls Based on Local Maximum Synchronous Extraction of Generalized S-Transforms

    by Mingchao Zhu, Rui Feng, Xiaofeng Zhang, Pengsheng Li, Binghua Su

    Bowhead whales (Balaena mysticetus) are an important species in the Arctic ecosystem, but their conservation is challenged by environmental noise from shipping and climate change. Effective monitoring of Bowhead whale vocalizations is essential for their conservation, yet traditional acoustic methods face limitations in detecting weak and non-stationary signals amidst complex background noise. In this study, we propose a novel method, Local Maximum Simultaneous Extraction of Generalized S-Transforms (LMSEGST), to enhance the feature extraction of Bowhead whale calls. The LMSEGST method integrates generalized S-transforms with local maximum extraction, improving time-frequency resolution and noise immunity. We compare the performance of LMSEGST with traditional methods (STFT, GST, and LMSST) using synthetic and real-world datasets. The results show that LMSEGST outperforms the other methods, with a 28.32% reduction in Rayleigh entropy compared to STFT at 5 dB SNR and a 28.24% reduction compared to LMSST at 10 dB SNR. Additionally, LMSEGST maintained higher SNR values, demonstrating superior noise resistance. These findings suggest that LMSEGST offers a more robust solution for acoustic monitoring of Bowhead whales, particularly in noisy, Arctic environments, contributing to more effective conservation strategies for this species.


  • Open Access

    SE-MTCAELoc: SE-Aided Multi-Task Convolutional Autoencoder for Indoor Localization with Wi-Fi

    by Yongfeng Li, Juan Huang, Binghua Su

    Indoor localization finds wide-ranging applications in user navigation and intelligent building systems. Nevertheless, signal interference within complex indoor environments and challenges regarding localization generalization in multi - building and multi - floor scenarios have restricted the performance of traditional positioning methods based on WiFi fingerprinting. To tackle these issues, this paper presents the SE - MTCAELoc model, a multi - task convolutional autoencoder approach that integrates a squeeze - excitation (SE) attention mechanism for indoor positioning. Firstly, the method pre - processes WiFi Received Signal Strength (RSSI) data. In the UJIIndoorLoc dataset, the 520 - dimensional RSSI features are extended to 576 dimensions and reshaped into a 24×24 matrix. Meanwhile, Gaussian noise is introduced to enhance the robustness of the data. Subsequently, an integrated SE module combined with a convolutional autoencoder (CAE) is constructed. This module aggregates channel spatial information through squeezing operations and learns channel weights via excitation operations. It dynamically enhances key positioning features and suppresses noise. Finally, a multi - task learning architecture based on the SE - CAE encoder is established to jointly optimize building classification, floor classification, and coordinate regression tasks. Priority balancing is achieved using weighted losses (0.1 for building classification, 0.2 for floor classification, and 0.7 for coordinate regression). Experimental results on the UJIIndoorLoc dataset indicate that the accuracy of building classification reaches 99.57%, the accuracy of floor classification is 98.57%, and the mean absolute error (MAE) for coordinate regression is 5.23 meters. On the TUT2018 dataset, the floor classification accuracy attains 98.13%, with an MAE of 6.16 meters. These results suggest that the SE - MTCAELoc model can effectively enhance the localization accuracy and generalization ability in complex indoor scenarios and meet the localization requirements of multiple scenarios.

  • Spcformer: spatial perception correction transformer for semantic segmentation of scene parsing

    by Zhengan Lu, Zhuang Zhou, Shuobin Wei, Zizhao Yuan, Binghua Su

  • Open Access

    Zoom Auxiliary Imaging Lens Design for a Modulation Transfer Function Test System

    by Yicheng Sheng, Sihan Xu, Caishi Zhang, Binghua Su, Dingxiang Cao, Zhe Chen

    In this paper, we propose a zoom auxiliary imaging lens based on the four-component mechanical zoom method for a modulation transfer function (MTF) test system. The auxiliary imaging lenses of the current MTF test system typically use fixed-focus optical systems, which are unable to meet the test scenarios of fast and batch measurement and measure lenses with an extensive focal length range. Compared with the fixed-focus auxiliary imaging lens, the zoom auxiliary imaging lens can simultaneously satisfy the measurement of wide-angle and telephoto miniature lenses without losing measurement accuracy. The entrance pupil distance of the zoom lens is greater than that of traditional lenses, and it is constant for each focal length of the zoom lens. The zoom lens uses an intermediate real image surface to obtain the perfect image quality and lower the diameter of the rear group. Additionally, the zoom lens dynamically adjusts magnification to optimize image size and align with the detector’s pixel resolution, thereby preventing undersampling and enhancing measurement precision. The optical design is optimized for stability, delivering high resolution and minimal aberrations across the zoom range. The image quality of the zoom lens is nearly at the diffraction limit at each focal length, which significantly reduces the impact of the auxiliary lens on MTF test results, enhancing both flexibility and accuracy. This design is particularly well suited for testing miniature lenses in optoelectronic technology applications.

  • HDBFormer: Efficient RGB-D Semantic Segmentation With a Heterogeneous Dual-Branch Framework

    by Shuobin Wei, Zhuang Zhou, Zhengan Lu, Zizhao Yuan, Binghua Su

    In RGB-D semantic segmentation for indoor scenes, a key challenge is effectively integrating the rich color information from RGB images with the spatial distance information from depth images. However, most existing methods overlook the inherent differences in how RGB and depth images express information. Properly distinguishing the processing of RGB and depth images is essential to fully exploiting their unique and significant characteristics. To address this, we propose a novel heterogeneous dual-branch framework called HDBFormer, specifically designed to handle these modality differences. For RGB images, which contain rich detail, we employ both a basic and detail encoder to extract local and global features. For the simpler depth images, we propose LDFormer, a lightweight hierarchical encoder that efficiently extracts depth features with fewer parameters. Additionally, we introduce the Modality Information Interaction Module (MIIM), which combines transformers with large kernel convolutions to interact global and local information across modalities efficiently. Extensive experiments show that HDBFormer achieves state-of-the-art performance on the NYUDepthv2 and SUN-RGBD datasets.


  • Effects of Different Normalization on the ESRGAN

    by Yongqi Tian, Jialin Tang, Lihong Niu, Binghua Su, Yulei An

  • Open Access

    Multi-Integration Time Adaptive Selection Method for Superframe High-Dynamic-Range Infrared Imaging Based on Grayscale Information

    by Xingyu Tao, Weiqi Jin, Jianguo Yang, Shuo Li, Binghua Su, Minghe Wang

    With the development of superframe high-dynamic-range infrared imaging technology that extends the dynamic range of thermal imaging systems, a key issue that has arisen is how to choose different integration times to obtain an HDR fusion image that contains more information. This paper proposes a multi-integration time adaptive method, in order to address the lack of objective evaluation methods for the selection of superframe infrared images, consisting of the following steps: image evaluation indicators are used to obtain the best global exposure image (the optimal integration time); images are segmented by region-growing point to obtain the ambient/high-temperature regions, selecting the local optimum images with grayscale closest to the medium grayscale of the IR imaging system for the two respective regions (lowest and highest integration time); finally, the three images above are fused and enhanced to achieve HDR infrared imaging. By comparing this method with some existing integration time selection methods and applying the proposed method to some typical fusion methods, via subjective and objective evaluation, the proposed method is shown to have obvious advantages over existing algorithms, and it can optimally select the images from different integration time series images to form the best combination that contains full image information, expanding the dynamic range of the IR imaging system.


  • GIU-GANs: Global Information Utilization for Generative Adversarial Networks

    by Yongqi Tian, Xueyuan Gong, Jialin Tang, Binghua Su, Xiaoxiang Liu, Xinyuan Zhang


  • Open Access

    GIU-GANs: Global Information Utilization for Generative Adversarial Networks

    by Yongqi Tian, Xueyuan Gong, Jialin Tang, Binghua Su, Xiaoxiang Liu, Xinyuan Zhang

    In recent years, with the rapid development of artificial intelligence, image generation based on deep learning has dramatically advanced. Image generation based on Generative Adversarial Networks (GANs) is a promising study. However, since convolutions are limited by spatial-agnostic and channel-specific, features extracted by traditional GANs based on convolution are constrained. Therefore, GANs are unable to capture any more details per image. On the other hand, straightforwardly stacking of convolutions causes too many parameters and layers in GANs, which will lead to a high risk of overfitting. To overcome the aforementioned limitations, in this paper, we propose a new GANs called Involution Generative Adversarial Networks (GIU-GANs). GIU-GANs leverages a brand new module called the Global Information Utilization (GIU) module, which integrates Squeeze-and-Excitation Networks (SENet) and involution to focus on global information by channel attention mechanism, leading to a higher quality of generated images. Meanwhile, Batch Normalization(BN) inevitably ignores the representation differences among noise sampled by the generator, and thus degrade the generated image quality. Thus we introduce Representative Batch Normalization(RBN) to the GANs architecture for this issue. The CIFAR-10 and CelebA datasets are employed to demonstrate the effectiveness of our proposed model. A large number of experiments prove that our model achieves state-of-the-art competitive performance.