Name: RAMÓN GIOSTRI CAMPOS
Publication date: 20/10/2021
Advisor:
Name |
Role |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Advisor * |
| FLAVIO GARCIA PEREIRA | External Examiner * |
| KLAUS FABIAN COCO | External Examiner * |
Summary: This thesis addresses on some aspects of the deconvolution of radio astronomical images captured by the aperture synthesis process, an interferometric process in which the image is captured in the space of frequencies and must be restored in the luminous space. This image
restoration is a challenging inverse problem given the characteristics of the capture process. Deconvolution is the final step in the process that generates the reconstructed image, and for the last 50 years, the radio astronomical community has been using algorithms from the CLEAN family, either individually or in association with other deconvolution methods. Even the most modern versions of the CLEAN algorithms have features that is undesirable for the scientific community. In this context, this thesis addresses two of these characteristics: the deconvolution of negative luminosity, physically impossible, and the high number of human choices necessary for the algorithm to work. An improvement on the multi-scale CLEAN algorithm (MS-CLEAN) were proposed to enable monitoring using Shannon entropy so that scales of little interest were removed from the search space, while avoiding the arising of negative brightness. The proposed algorithm was called Relevant Component CLEAN (RC-CLEAN), which proved to be up to 4 times faster than MS-CLEAN without prejudice to structure identification and noise reduction. The objective performance evaluation was carried out using the metrics SSIM and P SNR.
For simulated data, it is obtained the same quality for the SSIM and gains up to 11dB in the RC-CLEAN P SNR. RC-CLEAN also presents results similar to that obtained by computational tools used by large astronomical laboratories dealing with real data, being as competitive as the most relevant algorithms in this area. This research proposed a metric that objectively assesses the deconvolution processes of radio interferometric images, that is, it is a problem in the field of Image Quality Assessment (IQA). This area has a significant lack of study since objective metrics dedicated to astronomical radio images were proposed over 30 years ago and have undergone little or no change since then.
There is particular interest in developing a no-reference metric (NR) since ground truth is generally not available for real data. The metric is called Restrored-Residual-Image-Metric (RRIM), being inspired by SSIM as this metric simulates human perception, and has the flexibility to emphasize aspects deemed most relevant in the images astronomical radio. The metric RRIM presented a competitive behavior against the traditional radio astronomy NR metric, the Dynamic Range (DR). In fact, in most cases, RRIM had correlation coefficients higher than DR in comparisons made against the metrics SSIM
and P SNR. In cases WHERE it was not greater than DR, RRIM was equivalent.
Keywords: Image restoration, Radio Interferometric Images, Deconvolution, CLEAN, Image Quality Assessment.
