Mingyang Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mingyang Li has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Wei Zhang
School of Microelectronics, Tianjin University, Tianjin 300072, China
4 shared publications - Changsong Liu
School of Microelectronics, Tianjin University, Tianjin 300072, China
3 shared publications Yimeng Fan
School of Microelectronics, Tianjin University, Tianjin, CHINA
3 shared publicationsLixue Xu
School of Microelectronics, Tianjin University, Tianjin, CHINA
2 shared publications- Yanyan Liu
College of Electronic Information and Optical Engineering, Nankai University, Tianjin, China
2 shared publications - Yimeng Fan
School of Microelectronics Tianjin University China
2 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 9147-9156
SpikeGeo: Dual-Guidance Enhanced Spiking Neural Network for Energy-Efficient Cross-View Geo-Localization
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 9147-9156
Open AccessSpikeGeo: Dual-Guidance Enhanced Spiking Neural Network for Energy-Efficient Cross-View Geo-Localization
- Open Access
- 13 October 2026
- Published by Elsevier in SSRN Electronic Journal
Efficient Single-Pixel Imaging Reconstruction via Physics-Guided Wavelet–Fourier Dual-Domain Learning
Single-pixel imaging (SPI) records the integrated intensity of an encoded scene with a single-pixel detector and reconstructs a two-dimensional image from undersampled one-dimensional measurements. Because SPI measurements mix full-image information, global structures are difficult to disentangle at low sampling ratios, while low-energy high-frequency components, such as edges and directional textures, are prone to degradation. Transformers improve global modeling, but the quadratic complexity of self-attention limits lightweight deployment. Existing local-enhancement and multi-scale methods improve detail recovery, yet still struggle to balance global structures, directional high-frequency information, and deployment efficiency at low sampling ratios. To address these issues, we propose DuFRe-Net, a physics-guided dual-frequency-domain reconstruction network. It unfolds the Kronecker SPI model and embeds measurement-consistency constraints into each reconstruction stage. A Dual-Domain Wavelet–Fourier Modulation module is developed. The Fourier branch establishes long-range structural relationships at low cost, while the wavelet branch uses stable low-frequency structures to guide directional high-frequency recovery. Both branches adopt lightweight multi-head modulation to avoid explicit correlation computation. The Partial-Channel Feed-forward Network further reduces redundancy in feature refinement. Experiments across multiple sampling ratios on Set11 and BSD68, together with validation on a self-built real SPI system, show that DuFRe-Net achieves the best or competitive reconstruction performance. At SR=0.10, it contains only 0.645 M parameters and occupies 2.46 MB. These results demonstrate a favorable balance among global structure recovery, directional detail preservation, and deployment efficiency at low complexity.
- 13 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEfficient Single-Pixel Imaging Reconstruction via Physics-Guided Wavelet–Fourier Dual-Domain Learning
Single-pixel imaging (SPI) records the integrated intensity of an encoded scene with a single-pixel detector and reconstructs a two-dimensional image from undersampled one-dimensional measurements. Because SPI measurements mix full-image information, global structures are difficult to disentangle at low sampling ratios, while low-energy high-frequency components, such as edges and directional textures, are prone to degradation. Transformers improve global modeling, but the quadratic complexity of self-attention limits lightweight deployment. Existing local-enhancement and multi-scale methods improve detail recovery, yet still struggle to balance global structures, directional high-frequency information, and deployment efficiency at low sampling ratios. To address these issues, we propose DuFRe-Net, a physics-guided dual-frequency-domain reconstruction network. It unfolds the Kronecker SPI model and embeds measurement-consistency constraints into each reconstruction stage. A Dual-Domain Wavelet–Fourier Modulation module is developed. The Fourier branch establishes long-range structural relationships at low cost, while the wavelet branch uses stable low-frequency structures to guide directional high-frequency recovery. Both branches adopt lightweight multi-head modulation to avoid explicit correlation computation. The Partial-Channel Feed-forward Network further reduces redundancy in feature refinement. Experiments across multiple sampling ratios on Set11 and BSD68, together with validation on a self-built real SPI system, show that DuFRe-Net achieves the best or competitive reconstruction performance. At SR=0.10, it contains only 0.645 M parameters and occupies 2.46 MB. These results demonstrate a favorable balance among global structure recovery, directional detail preservation, and deployment efficiency at low complexity.
- 30 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 5621717
Frequency-Guided Mutual Distillation for Lightweight Cross-Modal Remote Sensing Image Retrieval
Cross-modal remote sensing image retrieval (CMR-SIR) is vital for Earth observation but remains challenged by the fundamental heterogeneity gap. Existing hashing methods often struggle to reconcile fine-grained representation with semantic consistency, primarily due to resolution-induced physical degradation and spatial–spectral discrepancies. To address these issues, we propose the Frequency-Guided Mutual Distillation Network (FG-MD), a lightweight framework designed for efficient and consistent feature alignment. The proposed approach first introduces a fine-grained refinement stage featuring a Frequency Domain Full-Spectrum Perceiver (FDFSP) module and an Adaptive Dynamic Filter (ADF), which synergistically enhance high-frequency structural details and suppress environmental clutter. Subsequently, a Dual-Attention Mutual Distillation (DAMD) strategy is devised to enforce spatial-semantic consistency via symmetric collaborative learning without incurring additional inference overhead. Extensive experiments conducted across multiple benchmarks demonstrate that FG-MD consistently achieves state-of-the-art performance with only 22.05 M parameters, offering an optimal efficiency-accuracy trade-off across all evaluated datasets. Furthermore, to validate generalization, we construct the Indoor Geometric Structure (IGS) benchmark for GPS-denied UAV reconnaissance. This dataset constitutes a challenging isomorphic extension of remote sensing tasks, where FG-MD significantly outperforms competitive baselines under severe viewpoint variations.
- 30 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 5621717
Frequency-Guided Mutual Distillation for Lightweight Cross-Modal Remote Sensing Image Retrieval
Cross-modal remote sensing image retrieval (CMR-SIR) is vital for Earth observation but remains challenged by the fundamental heterogeneity gap. Existing hashing methods often struggle to reconcile fine-grained representation with semantic consistency, primarily due to resolution-induced physical degradation and spatial–spectral discrepancies. To address these issues, we propose the Frequency-Guided Mutual Distillation Network (FG-MD), a lightweight framework designed for efficient and consistent feature alignment. The proposed approach first introduces a fine-grained refinement stage featuring a Frequency Domain Full-Spectrum Perceiver (FDFSP) module and an Adaptive Dynamic Filter (ADF), which synergistically enhance high-frequency structural details and suppress environmental clutter. Subsequently, a Dual-Attention Mutual Distillation (DAMD) strategy is devised to enforce spatial-semantic consistency via symmetric collaborative learning without incurring additional inference overhead. Extensive experiments conducted across multiple benchmarks demonstrate that FG-MD consistently achieves state-of-the-art performance with only 22.05 M parameters, offering an optimal efficiency-accuracy trade-off across all evaluated datasets. Furthermore, to validate generalization, we construct the Indoor Geometric Structure (IGS) benchmark for GPS-denied UAV reconnaissance. This dataset constitutes a challenging isomorphic extension of remote sensing tasks, where FG-MD significantly outperforms competitive baselines under severe viewpoint variations.
- 16 December 2025
- Published by Elsevier in Pattern Recognition
- Vol. 174
MS2Edge: Towards energy-efficient and crisp edge detection with multi-scale residual learning in SNNs
- 16 December 2025
- Published by Elsevier in Pattern Recognition
- Vol. 174
MS2Edge: Towards energy-efficient and crisp edge detection with multi-scale residual learning in SNNs