Image restoration model training method, image restoration method and device

Abstract

The invention discloses an image restoration model training method, an image restoration method and device, and the image restoration method comprises the steps: carrying out the mapping of a first image set with first image quality through a bidirectional mapping generative adversarial network model, and obtaining a second image set with second image quality; inputting the first image set and thesecond image set into a depth generation model, and performing image restoration on the second image set through the depth generation model to obtain a restored third image set; determining whether to retrain the depth generation model according to the loss function, and determining the trained depth generation model as an image restoration model; wherein the loss function is determined accordingto the image feature information of the first image set, the image feature information of the third image set and the discrimination result of the discriminator. By means of the mode, the working efficiency can be improved, and the application range of the image restoration model is widened.

Description

technical field [0001] The present application relates to the technical field of image recognition, in particular to a training method for an image restoration model, an image restoration method and a device. Background technique [0002] With the development of technology, users often need to upload some picture information. For example, when a user applies for a financial loan or other identification scenarios online, the user is required to upload a personal ID photo. In order to reduce the network delay, it is often necessary to perform lossy compression on the ID photo uploaded by the user for transmission; at the same time, due to the influence of factors such as the user's shooting device and the ambient light of the scene, this will lead to uneven quality of the photos we get. Some of the photos may even be of poor quality, which requires image restoration. [0003] Currently, deep generative models are used to repair images of ID card photos. Training deep generat...

AI Technical Summary

Problems solved by technology

[0004] This artificial simulation method is time-consuming and labor-intensive, and requires constant trial and adjustment. Sometimes it is difficult to simulate similarly, and it is also difficult to fully reproduce the lighting and shooting effects in various complex scenes.

However, when such a paired training set is used to train a deep classification model, the image restoration effect of the deep classification model will be poor.

Images