Image restoration model, method and device based on complex task decomposition regularization

Abstract

The embodiment of the invention provides an image restoration model, method and device based on complex task decomposition regularization. The image restoration model comprises a feature extraction and denoising subnet, a deblurring subnet and an image reconstruction subnet. The feature extraction and denoising subnet comprises a convolutional neural network, a deep convolutional neural network and a denoising network; wherein the convolutional neural network is used for carrying out feature extraction on an original blurred image to obtain a first feature map; the deep convolutional neural network is used for performing feature extraction on the original blurred image to obtain a second feature map; the noise reduction network is used for reconstructing a blurred image feature map according to the first feature map and the second feature map; wherein the deblurring subnet is used for removing turbulence blur in the blurred image feature map to obtain a third feature map; the image reconstruction subnet is used for carrying out image reconstruction according to the third feature map and outputting a reconstructed clear image; by decomposing the task, the complexity of the problem is reduced, the network generalization ability is enhanced, the over-fitting problem is reduced, and the image restoration effect is improved.

Description

technical field [0001] The present application relates to the technical field of image restoration, in particular, to an image restoration model, method and device based on complex task decomposition regularization. Background technique [0002] Traditionally, in order to solve complex primary vision tasks, an easy-to-think method is to increase the complexity of the model and improve the representation ability of the model. A more complex model means more parameters, making the model prone to overfitting. In order to solve the problem of neural network over-fitting, people have done a lot of research and thinking. Generalization bounds have been characterized for many functions, and many of these limits are obtained by some form of regularization (usually L2 regularization) or by limiting the complexity of the class of functions. In addition to the above methods, mainstream strategies to prevent overfitting during neural network training have recently proposed network pre...

AI Technical Summary

Problems solved by technology

[0003] However, low-level visual complex reconstruction tasks, especially the restoration of multi-factor interleaved degraded images, are very difficult

Compared to single-factor degraded images (e.g., motion blur), the restoration is more difficult, the degradation pattern is harder to learn, and the problem space is larger

The traditional end-to-end deep neural network image restoration method is difficult to learn the degraded inverse transformation mode of the image under the condition of limited training data learning, so the restoration effect is not good

Images