Image super-resolution algorithm based on multi-scale information distillation
To address the challenge of deploying image super-resolution algorithms based on deep learning on edge devices such as satellite navigation systems,due to the complexity of convolutional neural network frameworks and extensive computations,a novel multi-scale information distillation network(MSIDN)was proposed to reconstruct super-resolution images.This network utilizes a multi-stage strategy to progressively restore high-quality super-resolution images,with each stage composed of multi-scale information distillation coding-decode modules(MIDCB).In the encoding phase,MIDCB performs channel-wise split encoding on feature channels to retain shallow information and extract effective high-frequency signals.In the decoding phase,it enhances high-frequency signals and employs channel attention to merge the coding-decoding features of split channels.MSIDN learns more discriminative high-frequency feature representations and structural content information from MIDCB,not only enhancing the reconstruction performance of the super-resolution network but also meeting lightweight network structures.Conducting 4x magnification experiments on four public datasets including Set5,Set14,BSD100,and Urban100,the results showed an increase of 0.89,0.02,0.01,and 0.34 dB in peak signal-to-noise ratio compared to the enhanced deep residual super-resolution algorithm,respectively.The reconstructed images exhibited superior content structure and edge textures compared to other mainstream super-resolution algorithms,demonstrating the superiority of MSIDN in single-image super-resolution reconstruction.
image super-resolutionconvolutional neural networkmulti-scale information distillationchannel attention
万方数据
维普
