多尺度深层特征蒸馏的图像超分辨率重建

打开文本图片集

关键词：图像超分辨率重建;卷积神经网络;轻量级；多尺度特征蒸馏;注意力机制中图分类号：TP391.4文献标识码：Adoi：10.37188/OPE.20253310.1657 CSTR：32169.14.OPE.20253310.1657

Image super-resolution reconstruction of multi-scale deep feature distillation

LI Xiang ⋅1，2 ， XIONG Ling 1，2* ，YE Daohui 3 ， LI Shufan³

（1. School of Artificial Intelligence and Automation，Wuhan University ofScience and Technology，Wuhan 430081， China;

2. Engineering Research Center for Metallurgical Automation and Measurement Technology of Ministry of Education，Wuhan University of Science and Technology，Wuhan 43OO8l，China;

3.Sinopec JiangDiamond Oil Machinery Co，Ltd，Wuhan 4302OO，China） * Corresponding author， E-mail： xiongling@wust. edu. cn

Abstract：Aiming at the problem that existing super-resolution reconstruction algorithms were difficult to fully utilize multi-scale information and deep features of images，an image super-resolution reconstruction method based on multi-scale deep feature distillation（MSDFDN） was proposed. First， ConvNeXt convo lution was used to replace traditional convolution layers，increasing network depth with lower computational cost to improve performance. Second，a multi-scale deep feature distillation module was designed.By constructing ConvNeXt convolution layers of different scales and combining them with a residual feature distilation mechanism，multi-scale deep features in residual blocks were extracted while bypasing rich low-frequency information. Finally，an attention mechanism was introduced at the end of the module to adaptively weight extracted features，enabling the network to focus more on high-frequency information. Compared with other advanced lightweight super-resolution reconstruction algorithms on benchmark datasets and the self-built PDC bit composite dataset，the peak signal-to-noise ratio and structural similarity quantitative metrics of images obtained by this method showed improvement. Especially on the Urban100 dataset with more detailed information，the peak signal-to-noise ratio of the four-fold reconstructed image reaches 26.49 dB，and the structural similarity reaches O.797 6. Experimental results show that the proposed method has better objective and subjective measurement results.

Key words： image super-resolution； convolutional neural network；lightweight；multi-scale feature distillation；attention mechanism

1引言

在当今数字图像处理和计算机视觉领域，单图像超分辨率重建（SingleImageSuper-Resolution，SISR）技术受到了广泛关注。（剩余18870字）