Neural network image restoration model generation method and device

The embodiment of the invention provides a neural network image restoration model generation method and device based on low information entropy and high tolerance, and the method comprises the steps: obtaining a training set and a test set, and setting the algorithm hyper-parameters of a neural network image restoration model; acquiring a natural image and a mask image, and performing preprocessing and data batching; inputting a neural network image restoration model, performing preliminary restoration through the neural network image restoration model to obtain a preliminary restoration image, performing wavelet transform-based component decomposition on the preliminary restoration image to obtain a main semantic component and a texture structure component, performing unlimited style modulation on the main semantic component to obtain a texture structure component; texture likelihood sparse representation reconstruction is carried out on texture structure components, wavelet inverse transformation-based comp