XSXiaofei Song

Xiaofei Song

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

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Xiaofei Song has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • DBGENet: a dual-branch global-context and edge-aware encoder network for photovoltaic panel segmentation under complex remote sensing scenes

    by Chen Xie, Mingju Chen, Xingzhong Xiong, Yuyan Wang, Yangming Luo, Xueyang Duan, Xiaofei Song, Jie Rao, Xingyue Zhang, Zhihao Lin

    Accurate segmentation of photovoltaic (PV) panels from remote sensing imagery is essential for power plant management, yet remains challenging due to complex backgrounds and high reflectivity. To address this, we propose DBGENet, a unified network integrating global context modeling, local detail preservation, and edge-aware learning. Specifically, a dual-branch encoder collaboratively extracts global semantics and fine-grained spatial features. This architecture is enhanced by a global channel-spatial attention module to suppress background interference, and an edge-aware lightweight bottleneck module to incorporate explicit boundary priors for improved contour integrity. Extensive experiments on the publicly available Earth System Science Data (ESSD) benchmark, which comprises UAV (PV01, 0.1 m), aerial (PV03, 0.3 m), and satellite (PV08, 0.8 m) imagery, demonstrate DBGENet’s superiority. Our method achieves an mIoU of 94.48% for rooftop PV01, 96.63% for ground-mounted PV03 scenarios such as water surfaces, 80.13% for satellite PV08, exhibiting exceptional boundary accuracy and robustness across diverse complex scenes.

  • FMSNet: Frequency-domain aware and multiscale edge enhancement network for remote sensing land cover image segmentation

    by Yangming Luo, Chen Mingju, Xingzhong Xiong, Xueyang Duan, Xiaofei Song, Zhihao Lin, Chen Xie

    Land cover segmentation in remote sensing images faces challenges such as drastic scale variations, loss of high-frequency details, and blurred boundaries. Existing methods struggle to achieve fine-grained local structure restoration while maintaining global semantic consistency. To address this, this paper proposes a frequency-domain-aware, multi-scale edge enhancement network, FMSNet. By synergistically fusing the multi-scale local modeling capabilities of CNNs, the global semantic representation capabilities of Transformers, and a frequency-domain explicit awareness mechanism, FMSNet achieves joint modeling of semantic and structural information in remote sensing images. In the encoding stage, a Directional Detail Aware Stem module is designed. Through collaborative modeling of Directional Depthically Separable Convolutions and Patch Embedding, the directional detail representation of initial features is enhanced with lightweight overhead. In the decoding stage, a Frequency-Domain-Edge Co-enhanced Decoder is constructed. This decoder utilizes four-branch multi-scale dilated convolutions to aggregate context, introduces frequency-domain attention at key layers to explicitly model high and low-frequency differences, and combines this with edge-aware adaptive upsampling to achieve fine-grained boundary reconstruction. Furthermore, frequency consistency loss and edge loss are introduced for joint constraints to promote the synergistic optimization of semantics, frequency domain, and boundary structure. On the Potsdam, LoveDA, and GID-6 datasets, FMSNet achieved mIoU of 88.76%, 54.88%, and 81.25%, respectively, which is significantly better than existing methods. It performs particularly well in the segmentation of small objects and complex textured features, verifying the effectiveness and generalization ability of the proposed method.

  • Open Access

    Theoretical Framework, Technical Evolution, and Future Prospects of Cross-Modal Mapping and Controllable Image Generation Under Multi-Source Heterogeneous Collaboration

    by Mingju Chen, Zhihao Lin, Xiaofei Song, Yangming Luo, Xueyang Duan, Senyuan Li, Chen Xie

    The rapid evolution of diffusion models has shifted visual synthesis from text-only inputs to precisely controlled generation driven by multi-source heterogeneous sensor signals (e.g., audio, 3D, and physiological data). This paper presents a systematic review of cross-modal mapping and controllable generation under multi-source collaboration. More precisely, we propose a unified “cross-modal mapping and injection” taxonomy by abstracting the intervention logic of heterogeneous signals. Fundamentally, we analyze these mechanisms in a backbone-agnostic manner, delineating the architectural transition from legacy U-Net dependencies to scalable architectures like Diffusion Transformers (DiTs) and tracing the technical evolution from single-source atomic driving to complex multi-source collaborative paradigms. Our mechanistic analysis reveals that seamless feature fusion heavily relies on gradient conflict resolution, rigorous arbitration, and dynamic disentanglement under multi-constraint scenarios. Furthermore, by systematizing current evaluation metrics, we identify intrinsic quality-controllability trade-offs through performance game analysis (e.g., Pareto optimization), yielding a scientifically grounded technical selection guide. The study concludes that overcoming current generation limitations necessitates integrating Hardware-in-the-Loop (HIL) deployment, PDE-driven physical constraints, and causal inference, laying the foundation for next-generation robust and real-time generative models.

  • DBHSE: a dual-branch heterogeneous and SSMSearch semantic-enhanced SNR-aware communication coding network

    by Xueyang Duan, Mingju Chen, Xingzhong Xiong, Xiaofei Song, Yangming Luo, Zhihao Lin, Chen Xie

    To address the issues of imbalanced semantic feature extraction, fragmented codeword matching, and insufficient noise resistance in wireless channels, this paper proposes a dual-branch heterogeneous and SSMSearch semantic-enhanced signal-to-noise ratio (SNR)-aware communication coding network (DBHSE), achieving complete preservation and reliable transmission of image semantics under bandwidth constraints and channel uncertainties. At the feature representation level, the network organically integrates the local texture capture advantages of ConvNeXt and the global modeling advantages of swing transformer through a multi-scale heterogeneous adaptive fusion framework. Combined with a semantic information entropy allocation mechanism, it utilizes a dynamic adaptive fusion module to strengthen core semantics and suppress redundancy from the source, laying a feature foundation for high-quality transmission. To address semantic biases during the encoding process, DBHSE constructs a codeword search module that integrates structured sequence modeling and state selection mechanisms. Through hierarchical precise matching and distortion correction, it promotes deep alignment between semantic features and the codebook. Furthermore, to address the uncertainty of the physical channel, the network further integrates an SNR-aware multi-level feature modulation framework, utilizing attention mechanisms and dynamic adaptation techniques to enhance channel awareness and ensure that critical information is not obscured in noisy environments. Experiments have shown that DBHSE can significantly improve image reconstruction quality under both additive white Gaussian noise (AWGN) and Rayleigh fading channels, achieving complete preservation and efficient transmission of semantic information in wireless environments.

  • Open Access

    MFCA-Transformer: Modulation Signal Recognition Based on Multidimensional Feature Fusion

    by Xiao Hu, Mingju Chen, Xingyue Zhang, Jie Rao, Senyuan Li, Xiaofei Song

    In order to solve the problems of modulation signals in low signal-to-noise ratio (SNR), such as poor feature extraction ability, strong dependence on single modal data, and insufficient recognition accuracy, this paper proposes a multi-dimensional feature MFCA-transformer recognition network that integrates phase, frequency and power information. The network uses Triple Dynamic Feature Fusion (TDFF) to fuse constellation, time-frequency, and power spectrum features through the adaptive dynamic mechanism to improve the quality of feature fusion. A Channel Prior Convolutional Attention (CPCA) module is introduced to solve the problem of insufficient information interaction between different channels in multi-dimensional feature recognition tasks, promote information transmission between various feature channels, and enhance the recognition ability of the model for complex features. The label smoothing technique is added to the loss function to reduce the overfitting of the model to the specific label and improve the generalization ability of the model by adjusting the distribution of the real label. Experiments show that the recognition accuracy of the proposed method is significantly improved on the public datasets, at high signal-to-noise ratios, the recognition accuracy can reach 93.2%, which is 3% to 14% higher than those of the existing deep learning recognition methods.


  • Open Access

    MFPI-Net: A Multi-Scale Feature Perception and Interaction Network for Semantic Segmentation of Urban Remote Sensing Images

    by Xiaofei Song, Mingju Chen, Jie Rao, Yangming Luo, Zhihao Lin, Xingyue Zhang, Senyuan Li, Xiao Hu

    To improve semantic segmentation performance for complex urban remote sensing images with multi-scale object distribution, class similarity, and small object omission, this paper proposes MFPI-Net, an encoder–decoder-based semantic segmentation network. It includes four core modules: a Swin Transformer backbone encoder, a diverse dilation rates attention shuffle decoder (DDRASD), a multi-scale convolutional feature enhancement module (MCFEM), and a cross-path residual fusion module (CPRFM). The Swin Transformer efficiently extracts multi-level global semantic features through its hierarchical structure and window attention mechanism. The DDRASD’s diverse dilation rates attention (DDRA) block combines convolutions with diverse dilation rates and channel-coordinate attention to enhance multi-scale contextual awareness, while Shuffle Block improves resolution via pixel rearrangement and avoids checkerboard artifacts. The MCFEM enhances local feature modeling through parallel multi-kernel convolutions, forming a complementary relationship with the Swin Transformer’s global perception capability. The CPRFM employs multi-branch convolutions and a residual multiplication–addition fusion mechanism to enhance interactions among multi-source features, thereby improving the recognition of small objects and similar categories. Experiments on the ISPRS Vaihingen and Potsdam datasets show that MFPI-Net outperforms mainstream methods, achieving 82.57% and 88.49% mIoU, validating its superior segmentation performance in urban remote sensing.