关键词: 图像处理, 深度学习, 超分辨率重建, 卷积神经网络, 跳跃连接

Abstract: Image super-resolution is a cheap and convenient method for image enhancement, which is of great
significance in the research on video surveillance, medical imaging, satellite remote sensing, and other fields. In this
paper, combined with the performance advantages of deep learning in image reconstruction, an enhanced residual
dense network (ERDN) was proposed. Firstly, the enhanced residual dense block with multiple convolution kernels
was employed to extract the details of the image. Then the multi-layer image information was filtered and reorganized
through the skip connection and feature multiplexing modules, so that the network model could make comprehensive
use of image information at different depths. Finally, the parameter vectors of the reconstructed model were
constrained, and the decoupling operation of length and orientation enabled the model to converge at a larger learning
rate to improve the model training speed. Experimental validation on multiple international public datasets shows that
the method achieves better subjective visual evaluation and objective quantitative evaluation. For example, for the
quadruple reconstruction task, the peak signal-to-noise ratio (PSNR) metric of the ERDN output image was improved
by 0.24 dB over the dense residual network (RDN) on the Urban100 dataset, and the number of model parameters was
reduced by about 50%, which can be applicable to super-resolution reconstruction of various scene images.

Key words: image processing, deep learning, super-resolution reconstruction, convolution neural network, skip
connection