Super-resolution reconstruction algorithm for JPEG compressed images

Study of super-resolution reconstruction algorithm for compressed images

Super-resolution reconstruction is a high-resolution image reconstructed by one or more low-resolution images, such as reconstructed resolution 0.25m resolution images from 4 1m resolution images, which are widely used in military/civil applications. The current super-resolution reconstruction method is divided into 3 categories: Based on interpolation, learning-based, reconstruction-based approach, has now been more research. But most of the algorithms are to study the normal image, and the research on the compression image/video Super-resolution reconstruction is very few. Recently reviewed some of the literature, and conducted a study, here to do some summary.

Related literature:

1. Super-resolution from compressed video

2, Bayesian Resolution enhancement of compressed Video

3, Robustweb Image/video super-resolution

4, self-learning-based single Image super-resolution

of a highly compressed Image

Algorithm principle:

Document 1/2: Document 1/2 is written by the same person, the author published a number of articles on the issue, the method is basically similar. The idea is similar to the conventional super-resolution method based on reconstruction, it is to establish the observation model, and establish the reconstruction cost function according to the data fidelity and regularization terms. The difference is that the data fidelity is somewhat different, because it is necessary to consider the quantization noise and the noise of the compressed image, rather than the usual assumed Gaussian noise, in addition to the regularization term, it is necessary to consider the block effect of JPEG image, so in addition to the general constraints, the high frequency of the block boundary is limited. See the literature for details.

Article 3: This document is relatively new, and it is 2010. The idea of the article is well understood, first the JPEG compression image to block the effect, and then super-resolution. In the phase of block effect, the author uses adaptive Pde regularization method, and proposes an adaptive method to determine the regularization intensity, while in the reconstruction phase, it is a sample-based method, which is relatively mature. In fact, this two-step approach is not reasonable, is not advocating such treatment, but should be put together to consider, as in document 1/2.

Article 4: This document uses a sparse dictionary-related method. For the image to be reconstructed, first the next sampling, and the original image and the lower sampling image are used Bm3d processing to extract the low frequency, subtract from the original image and the image of the high frequency (the author thinks that the block effect exists only in the high-frequency components, respectively, to obtain the original image and the lower sampling image of high frequency is to train the and a self-training dictionary). After detecting high frequency, the author proposes that the high frequency is divided into block effect block/block effect block, and the dictionary d1/d2 is established respectively. And for The Block effect dictionary D1, the author will its atomic clustering, that 1 parts of the atom is to contribute to the block effect, and the other part of the atom does not contain block effect, then for the input block, I only use the atom without block effect of reconstruction, to achieve the purpose of the block effect.

In fact, this way to block effect, the author wrote another article, efficient image/video deblocking VIA SPARSE representation, you can take a look. After looking at the better understanding of this article on the compression image super-resolution, is nothing more than the problem of dictionary construction.

For the time being so much, not well written, you crossing will see, there are problems can be discussed again!!!!

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Research interests: Image compression, super-resolution, image denoising, image restoration and other image quality improvement techniques

Super-resolution reconstruction algorithm for JPEG compressed images