YSYangang Sun

Yangang Sun

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Yangang Sun has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • SPRINT: Accelerating Mamba‐Based Remote‐Sensing Foundation Models With Two‐Stage Token Reduction and Texture‐Aware Reconstruction

    by Mingshuo Chen, Yiping Song, Haoke Zhan, Yangang Sun, Xiang Cheng

    Recent advances in state‐space models such as Mamba have opened a path toward linear complexity foundation models for ultrahigh‐resolution remote‐sensing (RS) imagery, yet training efficiency remains hindered by redundant background tokens and sequential autoregressive objectives. We introduce SPRINT (Sparse Pretraining with Reduced INput Tokens), the first Mamba–Transformer pretraining framework tailored to RS data that jointly addresses these limitations. SPRINT couples a hybrid Mamba–Transformer backbone with a masked autoregressive pretraining objective and two acceleration mechanisms, (i) State‐Preserving Token Condensation, which merges locally redundant tokens to compress background patches and (ii) attention‐guided token pruning, which leverages multihead self‐attention scores to keep globally informative tokens. On the decoder side, a texture‐aware partial reconstruction strategy reconstructs only the 25% masked tokens with the highest local binary pattern entropy, thereby focusing on learning texture‐rich foreground regions. Compared with the autoregressive RoMA baseline, SPRINT cuts GPU memory by 46% and increases pretraining efficiency by 1.67 while achieving superior transfer performance on downstream tasks such as scene classification, semantic segmentation, and object detection. Extensive ablations confirm that each proposed component contributes to the acceleration, establishing SPRINT an efficient route toward scalable RS foundation models that maintain competitive representation quality despite aggressive token sparsification.

  • Open Access

    MMW‐FC: A novel railway fastener detecting method based on millimetre wave radar for train positioning

    by Yangang Sun, Jinhai Li, Chaosan Yang, Zhankun Du, Jifeng Zhang, Xin Qiu

    A novel method is proposed for rail fastener detection based on millimetre‐wave (mmWave) radar, mmWave radar fastener counter (MMW‐FC), which can accurately detect and record the fasteners in real‐time as the train traverses its route. Under circumstances where GNSS signals remain unavailable for prolonged durations, precise train localisation can be accomplished by correlating the number of fasteners derived from this method with the corresponding track map. Initially, MMW‐FC utilises fast Fourier transform and adaptive beamforming to focus the energy reflected from fasteners. Subsequently, it applies an adaptive template‐matching algorithm to detect each fastener. Furthermore, by leveraging known fastener spacing and the average time for trains to pass adjacent fasteners, the Kalman filter can execute precise speed tracking, used as a speed reference when adjusting the matching template adaptively. The experimental results indicate that the proposed method can precisely count the fasteners the train encounters in diverse road and speed conditions. The fastener counter maintains the Counting Error less than 0.067%, the speed error stays below 1.8 km/h, and the maximum values of the mean absolute error and root mean square error for speed are 0.7337 and 0.9584 km/h, respectively.


  • High-Precision Railway Fastener Detecting via Deep Learning Radar for Intelligent Train Positioning

    by Yangang Sun, Jinhai Li, Chaosan Yang, Po Hu, Jifeng Zhang, Xin Qiu

    This paper introduces FC-Radio, a novel method for real-time detection of railway fasteners traversed utilizing millimeter-wave (MMW) radar, effectively determining a train’s relative position on the track. The technology can function independently from trackside equipment and seamlessly integrate with current train control systems. Integrating signal processing with deep learning, the core components of FC-Radio are the Fastener Reflection Extractor and the Deep Neural Network. The former employs adaptive beamforming to optimize the signal-to-clutter ratio (SCR) and utilizes periodic feature detection to pinpoint the spatial location of fasteners. The latter adopts a neural network architecture comprising Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory Networks (BiLSTM), and attention mechanisms for meticulous fastener detection. The experimental data were collected at the National Railway Test Center of China, encompassing a diverse range of train operational scenarios. Experimental results indicate that FC-Radio exhibits capability in the real-time and accurate detection of fasteners during the high-speed operation of trains. The Counting Error is remarkably low, at -0.029%, which equates to a distance error of 38.1 m over a total distance of 129.6 km.