摘要
针对热成像设备采集的红外图像纹理细节较粗导致去模糊效果不佳的问题,在基于深度学习的可见光图像去模糊网络基础上,提出了一种改进的适用于红外图像的单幅图像去模糊方法,该方法以多尺度U型网络为基础进行改进。首先,在编码模块中引入对偶注意力单元,增强网络的特征表达能力;其次,将快速傅里叶变换嵌入到特征融合模块,加强网络对于高低频信息的处理能力;最后,选择更优的激活函数和损失函数,实现更好的信息流动并提升模型的鲁棒性。在红外数据集上进行测试,并与原网络进行对比,结果表明,本文改进网络的图像细节恢复得更好,峰值信噪比提升了0.53 dB,红外图像去模糊效果良好。
For the poor deblurring effect due to coarse texture details of infrared images collected by thermal imaging equipment,an improved single-image deblurring method suitable for infrared images is proposed based on a deep learning-based visible light image deblurring network.The method is improved based on a multi-scale U-shaped network.First,a dual attention unit is introduced in the encoding module to enhance the feature representation ability of the network.Second,the fast Fourier transform is embedded into the feature fusion module to enhance the processing capability of the network for high and low frequency information.Finally,better activation function and loss function are selected to achieve better information flow and improve the robustness of the model.Testing on the infrared dataset and compared with the original network,the results show that the improved network in this study has achieved better restoration of image details,and the peak signal-to-nose ration is 0.53 dB higher than the original network.The infrared image has a good deblurring effect.
作者
张艳珠
赵赫
刘义杰
ZHANG Yanzhu;ZHAO He;LIU Yijie(Shenyang Ligong University,Shenyang 110159,China)
出处
《沈阳理工大学学报》
CAS
2024年第1期55-60,共6页
Journal of Shenyang Ligong University
基金
辽宁省教育厅高等学校基本科研项目(LJKZ0245)
装备预研重点实验室基金项目(2021JCJQLB055006)。