Instituto de Óptica “Daza de Valdés”

Spectral Pre-Adaptation for Restoring Real-World Blurred Images using Standard Deconvolution Methods

17 Aug, 2022 | Ciencias de la imagen-en

Do you want to see a demo of how the best image restoration algorithms work?

Classic image blur correction models are based on assumptions that are not real, such as offset equivalence and the circular boundary condition (CBC), which are rarely actually met in practice.
Offset equivalence means that the image has the same blur at all points and does not suffer from aliasing. The circular boundary condition assumes that the image is rectangular and repeats itself periodically, which would imply that the top of the image continues at the bottom and the end of the image on the left continues without sudden jumps to the right.
Furthermore, the standard procedure for evaluating the performance of image restoration algorithms is to use a set of test images, with known but artificial blurs, which gives performance values that sometimes do not agree with the results with real images.

Discrepancies between the simplified models and the actual fuzzy observations cause strong artifacts in image restoration.

What are artifacts

An artifact is a visible distortion in the treated image, such as visible lines that were not there before.

Últimas noticias

Photo of a woman wearing a hat with numerous artifacts on the edges and in the center of the image
In this photograph, artifacts are seen in the form of lines on the edges and in the form of texture in the rest of the image
The common remedy to avoid artifacts and improve performance is to increase the complexity of the correction algorithm in order to remove simplifying assumptions. However, this implies a lot of computational load for the restore task.

In this work the scientific team with the participation of Fraunhofer from Singapore, The Institute for Information Theory and Automation of the Czech Academy of Sciences and the Group of Image and Vision Sciences of the Institute of Optics has presented a spectral pre-adaptation method (called SPA) that pre-processes with the blurred images so that they can be restored using algorithms of lightweight standard deconvolution.

9 photos of a woman in a hat with different levels of blur and artifacts
Comparison of the new SPA algorithm with complex models
While state-of-the-art deconvolution methods are iterative, complex and non-linear, the spectral pre-fitting method is linear and quite simple.
The SPA pre-adaptation therefore serves as a union between the simple reconstruction methods and the real observations that have blurring, it makes the restoration more robust against artifacts, and also after applying SPA, the user is free to choose any method. restoration based on a standard defocus model.

The team has compared the usefulness of the algorithm with various real-life applications by conducting experiments. The results indicate that SPA, when combined with efficient restoration methods, greatly removes artifacts with lower computational cost compared to state-of-the-art restoration methods. More importantly, for some experiments we have also tested SPA on real images and the results are as good as simulations.
Our restoration recovers very blurred texts in the background, now becoming partially legible. An experiment shows that this can even be applied to photos taken not by professional cameras but by inexpensive commercial cell phones, even subjected to JPEG compression.

Two photos of a text, one is out of focus and in the other you can read part of the text
All this suggests that the proposed algorithm has great potential in its application to images captured in real life.

The algorithm can be tested on a website that has been prepared to publicize its capabilities.

Related News