Application of a Modified Generative Adversarial Network in the Superresolution Reconstruction of Ancient Murals

Algorithm 1

Algorithm training process.

	Input: low-resolution image and the corresponding high-resolution image
	Output: the generative and discriminant network models
(1)	Step 1: the low-resolution image in the dataset is read
(2)	Step 2: the features of the mural image are extracted, and then upsampling is performed to obtain the high-resolution image of the target size
(3)	Step 3: MSE is calculated to update the network model; the reconstructed high-resolution mural image is output
(4)	Step 4: steps 1–3 are repeated to optimize the network model until the MSE tends to be relatively stable
(5)	Step 5: the generated high-resolution image as the false sample, and the corresponding real high-resolution image as the real sample are input into the discriminant network
(6)	Step 6: high-resolution image features are extracted, and finally, a feature vector is output after the fully connected layer
(7)	Step 7: the sigmoid function is used to transform the feature vector into a probability value and then determine whether the input image is a real superresolution image
(8)	Step 8: the sum of the content loss value and countermeasure network loss value is summed, and the generative network model and the discriminant network model are updated and saved
(9)	Step 9: steps 4–8 are repeated to update and optimize the parameters of the generative network and discriminant network model until the loss value of the model tends to be stable and remains for a period of time