Research on Image Denoising and Super-Resolution Reconstruction Technology of Multiscale-Fusion Images

Table 1

Comparison of image denoising methods.

Basic network	Methods involved	Advantages	Disadvantages	Applicable noise
Denoising method based on convolutional neural network	N2N [21]	No pairwise training samples are needed, which overcomes the problem of insufficient pairwise training samples in real images	Shallow pixel level information utilization is low and texture details are easily lost	Real noise
	VDN [22]			Complex noise
	FFDNet [23]			Complex noise
	CBDNet [24]			Real noise
	PRIDNet [25]			Real noise
Denoising method based on residual network	FC-AIDE [26]	The problems of gradient disappearing and gradient explosion are solved effectively, and the convergence speed is accelerated	The use of dense connections leads to overfitting of networks, which affects the consistency between objective evaluation indexes and subjective visual effects	Complex noise
	CycleISP [27]			Real noise
	PANet [28]			Complex noise
	GRDN [29]			Real noise
Denoising method based on Generative Adversarial Network	GCBD [24]	It can generate realistic noise images, expand the real image dataset, and solve the problem of insufficient training samples	There are some problems such as unstable network training, slow convergence speed, and uncontrollable model	Real noise
	ADGAN [30]			Real noise
Denoising method based on Graph Neural Network	GCDN [31]	Complex noise distribution can be well fitted by the topology of graph network	Unstable dynamic topology will reduce the ability of feature expression	Real noise
	DeepGLR [32]			Real noise
	OverNet [33]			Complex noise