Three-sheet simple lens image restoration method based on convolution neural network CNN

Abstract

The invention discloses a three-sheet simple lens image restoration method based on a convolution neural network CNN. The method comprises steps of firstly, generating a blurred image and clear imagedata set corresponding to three sheets of simple lenses; constructing a convolution neural network CNN model used for end-to-end image restoration; by use of the generated data set to train the CNN model; and for newly shot blurred images, by use of the trained CNN model, directly obtaining restored clear images. According to the invention, a large number of optimization iteration processes of blind convolution and non-blind convolution image restoration in the current method can be avoided; there is no need to independently estimate the PSF of the simple lenses, so the image restoration process of the three sheets of simple lenses are quite simple and convenient; the image processing speed is fast; and the method is very important in simple lens calculation and imaging field.

Description

technical field [0001] The invention relates to the field of image restoration, in particular to a three-piece simple lens image restoration method based on a convolutional neural network (CNN). Background technique [0002] In recent years, computational imaging with simple lenses has gradually become a new research direction in the field of image restoration. Simple lens calculation imaging aims to use the optical lens structure as simple as possible at the front end combined with the post-image calculation method to obtain imaging quality similar to that of high-end cameras such as SLR cameras. Simple lens computational imaging can greatly reduce the cost of optical design of the lens, and has important research value in the fields of image restoration and optical design. [0003] The lens of a simple lens usually only contains one, two or three lenses, and the front optical structure is simple, so the aberration and chromatic aberration of the lens itself will cause the...

AI Technical Summary

Problems solved by technology

The main disadvantages of the existing methods are: (1) Both blind convolution and non-blind convolution image restoration methods require a large number of iterative optimization processes, and the image processing and calculation time is long; (2) in order to improve the accuracy of image restoration, it is generally necessary to estimate The PSF of the single lens is obtained, and then the image restoration process is performed separately. The entire image processing process is relatively complicated; (3) Even if the PSF has been estimated, each blurred image still needs to be processed by a non-blind convolution algorithm, which is not conducive to simple lens calculation imaging. Handle a large number of images in practice

Images