网站访问量
今日访问量: 88
基于NTEG-Net的多阶段图像去雨网络
Multi-stage Image Rain Removal Network Based On NTEG-Net
宋建辉;曹祎男;刘晓阳; SONG Jianhui;CAO Yinan;LIU Xiaoyang;针对编码-解码器结构易丢失图像细节导致去雨结果模糊，而单阶段网络因深度有限导致全局语义信息不准确的问题，提出一个基于NTEG-Net的多阶段图像去雨网络。首先，前两个阶段使用基于U-Net改进的NTEG-Net编码-解码网络在不同尺度上学习雨纹的特征；其次，在第三阶段使用重新参数化的残差特征网络代替编码-解码器，生成具有高分辨率且信息丰富的背景特征；最后，为保证每阶段信息高效传递，在阶段间加入跨阶段特征传递模块实现去雨。实验结果表明，该网络相比MPRNet网络在数据集Rain100L、Rain100H、Rain800上峰值信噪比分别提高了2.09 dB、0.24 dB、0.27 dB,结构相似性分别提高了0.009、0.003、0.006,能有效去除雨滴并恢复图像背景细节。
In order to solve the problem that the encoder-decoder structure is easy to lose image details and the global semantic information is inaccurate due to the limited depth of the single-stage network, a multi-stage image rain removal network based on NTEG-Net was proposed.Firstly, in the first two stages, the improved NTEG-Net encoding-decoding network based on U-Net was used to learn the features of rain patterns at different scales.Secondly, in the third stage, the reparameterized residual feature network was used instead of the encoder-decoder to generate high-resolution and information-rich background features.Finally, in order to ensure the efficient transmission of information at each stage, a cross-stage feature transfer module was added between each stage to realize the transmission of information at different stages, and finally to realize rain removal.Experimental results show that compared with the original network, the peak signal-to-noise ratio of the proposed network on the datasets Rain100L,Rain100H and Rain800 is increased by 2.09 dB,0.24 dB and 0.27 dB,respectively, and the structural similarity is increased by 0.009,0.003 and 0.006,respectively, which can effectively remove raindrops and restore the background details of the image.
数据受限场景下的无设备手势识别方法研究
Research on Device-Free Gesture Recognition Method in Data-Limited Scenarios
郭崇;陶硕;李鹤群; GUO Chong;TAO Shuo;LI Hequn;针对无设备手势识别(device free gesture recognition, DFGR)技术因环境差异导致的精度下降问题，提出了一种基于深度学习的跨场景手势识别模型ADAN。该模型的创新之处体现在三个方面：其一，采用改进的ResNet101架构构建特征提取器，通过优化卷积核配置，提升捕捉手势细微特征的能力；其二，设计一个基于注意力机制的特征选择模块，通过动态权重分配实现关键特征的过滤和强化；其三，采用改进的域对抗损失函数对不同样本应用不同的权重操作，实现不同环境下的特征对齐。为验证模型性能，本研究基于Widar 3.0数据集构建了六个典型的室内环境进行训练。实验结果表明，ADAN在跨场景手势识别任务中实现了84.41%的平均准确率，较传统方法提升了3.11个百分点，并优于现有基准模型，为解决基于无线感知手势识别中的场景迁移问题提供了新的技术方法。
To address the accuracy decline of device-free gesture recognition(DFGR) technology due to environmental differences, a cross-scenario gesture recognition model based on deep learning, named ADAN,is proposed.The innovation of this model lies in three aspects.Firstly, an improved ResNet101 architecture is adopted to construct the feature extractor, whichi enhances the ability to capture subtle gesture features by optimizing the convolutional kernel configuration.Secondly, a feature selection module based on the attention mechanism is designed, which filters and strengthens key features through dynamic weight allocation.Finally, an improved domain adversarial loss function is used to apply different weight operations to different samples, achieving feature alignment in different environments.To verify the performance of the model, six typical indoor environments for training are constructed in this study based on the Widar 3.0 dataset.The experimental results show that ADAN achieves an average accuracy of 84.415% in cross-scenario gesture recognition tasks, which is 3.115 percentage points higher than that employing traditional method and outperforms existing benchmark models, providing a new technical approach to solving the scene transfer problem in wireless perception-based gesture recognition.
公告栏
期刊信息
期刊名称： 沈阳理工大学学报（Journal of Shenyang Ligong University）
创办日期： 1982年
主管单位： 辽宁省教育厅
主办单位： 沈阳理工大学
出版单位：《沈阳理工大学学报》编辑部
刊期： 双月刊
电话： 024-24686097
Email： sgxb6097@sylu.edu.cn
国内统一刊号(CN)： CN 21-1594/T
国际标准刊号(ISSN)：ISSN 1003-1251