Super-resolution(SR)reconstruction addresses the challenge of enhancing image resolution,which is critical in domains such as medical imaging,remote sensing,and computational photography.High-quality image reconstruct...Super-resolution(SR)reconstruction addresses the challenge of enhancing image resolution,which is critical in domains such as medical imaging,remote sensing,and computational photography.High-quality image reconstruction is essential for enhancing visual details and improving the accuracy of subsequent tasks.Traditional methods,including interpolation techniques and basic CNNs,often fail to recover fine textures and detailed structures,particularly in complex or high-frequency regions.In this paper,we present Deep Supervised Swin Transformer U-Net(DSSTU-Net),a novel architecture designed to improve image SR by integrating Residual Swin Transformer Blocks(RSTB)and Deep Supervision(DS)mechanisms into the U-Net framework.DSSTU-Net leverages the Swin Transformer’s multi-scale attention capabilities for robust feature extraction,while DS at various stages of the network ensures better gradient propagation and refined feature learning.The ST block introduces a hierarchical self-attention mechanism,allowing the model to capture both local and global context,which is crucial for high-quality SR tasks.Moreover,DS applied at multiple layers in the decoder enables direct supervision on intermediate feature maps,accelerating convergence and improving performance.The DSSTU-Net architecture was rigorously evaluated on the DIV2K,LSDIR,SET5,and SET14 datasets,demonstrating its superior ability to generate high-quality images.Furthermore,the potential applications of this model extend beyond image enhancement,with promising use cases in medical imaging,satellite imagery,and industrial inspection,where high-quality image reconstruction plays a crucial role in accurate diagnostics and operational efficiency.This work provides a reference method for future research on advanced image restoration techniques.展开更多
基金supported in part by the National Natural Science Foundation of China(62263006)2021 Director’s Fund of the Guangxi Key Laboratory for Automatic Detection Technology and Instruments(YQ21107)+2 种基金Guilin University of Electronic Technology Scientific Research Fund Project(UF24014Y)Innovation Project of Guangxi Graduate Education(YCSW2024336)Middle-aged and Young Teachers’Basic Ability Promotion Project of Guangxi(2021KY0802).
文摘Super-resolution(SR)reconstruction addresses the challenge of enhancing image resolution,which is critical in domains such as medical imaging,remote sensing,and computational photography.High-quality image reconstruction is essential for enhancing visual details and improving the accuracy of subsequent tasks.Traditional methods,including interpolation techniques and basic CNNs,often fail to recover fine textures and detailed structures,particularly in complex or high-frequency regions.In this paper,we present Deep Supervised Swin Transformer U-Net(DSSTU-Net),a novel architecture designed to improve image SR by integrating Residual Swin Transformer Blocks(RSTB)and Deep Supervision(DS)mechanisms into the U-Net framework.DSSTU-Net leverages the Swin Transformer’s multi-scale attention capabilities for robust feature extraction,while DS at various stages of the network ensures better gradient propagation and refined feature learning.The ST block introduces a hierarchical self-attention mechanism,allowing the model to capture both local and global context,which is crucial for high-quality SR tasks.Moreover,DS applied at multiple layers in the decoder enables direct supervision on intermediate feature maps,accelerating convergence and improving performance.The DSSTU-Net architecture was rigorously evaluated on the DIV2K,LSDIR,SET5,and SET14 datasets,demonstrating its superior ability to generate high-quality images.Furthermore,the potential applications of this model extend beyond image enhancement,with promising use cases in medical imaging,satellite imagery,and industrial inspection,where high-quality image reconstruction plays a crucial role in accurate diagnostics and operational efficiency.This work provides a reference method for future research on advanced image restoration techniques.
文摘矿井关闭后,其上覆岩层及地表会再次发生变形,影响建(构)筑物安全运营。由于关闭矿井缺乏监管,导致地表形变时空演化规律及预测预警模型研究不足。为此,提出了一种分布式目标雷达干涉测量(Distributed scatter Interferometric synthetic aperture radar,DS-InSAR)、时间域高通滤波(Temporal High Pass Filtering,THPF)以及长短期记忆网络(Long Short Term Memory Network,LSTM)相结合的关闭矿井地表形变预测模型。以98景Sentinel-1A升轨影像为数据源,首先利用DS-InSAR方法联合PS(Persistent Scatterer)和DS点获取徐州西部关闭矿井2019−11—2022−12的时序地表沉降信息;然后利用THPF分解原始沉降序列获取高频和低频序列沉降信息;之后采用LSTM完成高低频子序列的形变预测,将高低频子序列预测值叠加获取最终的预测结果。结果表明:DS-InSAR监测点密度分布均匀,监测结果水准实测形变间的决定系数达到0.95;相较于LSTM模型,THPF-LSTM模型在预测点位上的最大均方根误差(RMSE,Root Mean Square Error)为3.0,最大平均绝对误差(Mean Absolute Error,MAE)为2.4,最大校正决定系数(Adjusted R-square)为0.9,优于传统LSTM模型的4.5、3.9和0.6,模型综合预测精度提升20%以上,能够准确反映出关闭矿井地表形变的趋势和波动性,可有效提高短期内关闭矿井沉降预测精度,实现了关闭矿井地表形变监测和预测的一体化分析。
