Automatic overall constraint super-resolution image gradient reconstruction method

Abstract

According to the super-resolution image reconstruction method based on image super-resolution reconstruction and realizing overall constraint driving, the overall auto-approximation degree of the image block is introduced into the sparse representation coefficient space to serve as a regularization item to constrain solution of the sparse representation coefficient in the reconstruction model, the reconstruction error is reduced, the geometric structure of the image is maintained, and the image reconstruction efficiency is improved. And noise and sawtooth phenomena are reduced. In order to further optimize the reconstruction effect of the method on image edge details and improve the calculation speed of the method, an overall self-adaptive constraint super-resolution image reconstruction method is provided, and the gradient features of the image are used as priori knowledge to guide image blocks to perform classification reconstruction. And solving of the Euclidean distance between the image blocks is replaced by solving of the included angle of the gradient direction angle between the image blocks, so that the image reconstruction speed is improved. The PSNR and the SSIM of the improved method are greatly improved, the edge of the reconstructed image is clearer and sharpener, and the average reconstruction time is only equivalent to 65.8% of the average reconstruction time before improvement.

Description

technical field [0001] The present application relates to a super-resolution image gradient reconstruction method, in particular to an automatic overall constraint super-resolution image gradient reconstruction method, which belongs to the technical field of super-resolution image reconstruction. Background technique [0002] Currently, there are two mainstream methods to improve image resolution: the first is to improve the performance of hardware sensors; the second is through software. Regarding hardware equipment, in order to improve the imaging quality, one of the effective and feasible ways is to improve the production technology of the sensor, so that the size of the sensor array (pixel) becomes smaller or the size of the chip becomes larger. However, there will be a certain limit to the reduction of the size of the sensor array, and additional noise will be generated beyond a certain range. In addition, increasing the chip size will lead to an increase in capacitance...

AI Technical Summary

Problems solved by technology

[0011] First, the image reconstructed by interpolation in the prior art is not a true HR image, because these commonly used interpolation methods can only increase the number of image pixels, but cannot increase the effective high-frequency information. The reconstructed image lacks sufficient detail information, and has visual distortion effects such as irregular edges, ringing, and blurring. This type of method does not add additional effective image detail information, and it is difficult to introduce prior knowledge of the image. The reconstructed image The HR image is blurry and the definition is not high, which does not meet the actual needs

The existing technology based on the multi-frame reconstruction method needs to perform motion estimation and registration on multi-frame LR images, which brings difficulties to practical applications

The existing technology lacks an adaptive super-resolution image reconstruction method that can effectively improve the quality of the reconstructed image and increase the calculation speed

[0012] Second, the classic sparse representation super-resolution image reconstruction method in the existing technology has serious noise and aliasing, lacks the introduction of the overall self-approximation degree of the image block into the sparse representation coefficient space, and lacks a regularization term to constrain the reconstruction In the solution of the sparse representation coefficient in the structural model, the reconstruction error is large, the geometric structure of the image cannot be maintained, the measures for suppressing image noise are lacking, the texture structure of the image cannot be smoothed, the image has many jagged phenomena, and the reconstructed image is not clear enough and sparse. The stability of the representation is poor, the image has many defects, the overall reconstruction error is large, and the quality of the reconstructed image is poor

[0013] Third, in the reconstruction results of the overall constraint method proposed in this application, there are excessive smoothing and blurring of the image edge and the problem of a large amount of calculation. The edge blurring phenomenon is caused by the principle of the overall self-approximation degree of the image block , this method performs the weighted average of the overall image block, although it can reduce the error of sparse reconstruction, it will suppress the high-frequency information of the image, resulting in the loss of high-frequency detail features of the image, especially the edge part of the image with a large gradient change. The structure is not accurate enough, the lack of linking the gradient feature classification of the image block to the super-resolution reconstruction, and the lack of using the gradient feature of the image as prior knowledge to guide the image block to classify and reconstruct, which is not conducive to the edge part of the image with a large gradient change Accurate reconstruction, can not perform sparse representation and reconstruction stably, lacks the use of global constraints to further improve the entire reconstruction results, lacks the step of selecting the overall neighborhood of image blocks in the overall constraints, and lacks the use of gradient direction angles of image blocks To replace the solution of the Euclidean distance of the image block, the edge part is relatively blurred, and more overall image blocks need to be calculated, so the calculation amount is large, and the reconstruction speed is not fast enough

Images