Improved blind super-resolution reconstruction algorithm based on multi-image fuzzy kernel estimation

Abstract

The invention discloses an improved blind super-resolution reconstruction method based on multi-image blur kernel estimation. It mainly includes the following steps: blurring the first frame picture in the input low-resolution video frame to different degrees, obtaining two images of the same scene with different blur degrees, and obtaining two pictures with different blur degrees; using the above Get two pictures of the same scene with different degrees of blur, use a robust deconvolution algorithm to generate a rough blur kernel, and use this blur kernel to deblur all video frames, and get a set of processed Image sequence f ^k , as the input for subsequent processing; use the curvature difference operator to extract the spatial structure information, and then cluster it to obtain the regional spatial adaptive weighting coefficient, which is used to adaptively weight the full variation and non-local mean regularization terms ;Use the adaptive weighting coefficient obtained above to weight the regularization term, so as to determine the reconstruction cost function; use the gradient descent method to optimize the reconstruction cost function, in which the fuzzy kernel estimation is performed again in each iteration process, and deblurring is performed, Finally, an output high-resolution image sequence is obtained.

Description

technical field [0001] The invention relates to image super-resolution reconstruction technology, in particular to an improved blind super-resolution reconstruction algorithm based on multi-image fuzzy kernel estimation, which belongs to the field of digital images. Background technique [0002] As the carrier of visual information, images and videos are an important way for human beings to obtain and transmit information. Therefore, it is of great significance to study and process image and video information. With the development of information technology, people have higher and higher requirements for the received information, especially in the application fields of medicine, remote sensing, astronomy and video surveillance, etc., it is necessary to obtain high-resolution video. However, when actually capturing video, it is often affected by factors such as light aberration, undersampling, atmospheric disturbance, defocus, and system noise, and the spatial resolution of th...

AI Technical Summary

Problems solved by technology

The traditional blur kernel estimation based on boundary gradient changes needs to select the straight line border of two uniform bright and dark areas on the image as the edge. Since it is difficult to find enough strong edges in the input low-resolution image, the blur kernel estimation the problem becomes more difficult

In most current super-resolution reconstructio

Images