Comparación automática de calidad entre imágenes astronómicas reconstruidas utilizando técnicas de Deep Learning.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
En este trabajo se abordó el problema de la evaluación objetiva de la calidad en imágenes astro nómicas reconstruidas a partir de observaciones interferométricas, un proceso que tradicionalmente depende de la inspección visual de expertos y carece de metodologías reproducibles. Es por esto que se propuso el desarrollo de un sistema automático basado en técnicas de Deep Learning, capaz de comparar múltiples reconstrucciones provenientes de una misma observación y establecer un ranking de calidad sin necesidad de contar con una imagen de referencia. Con este propósito, se generó un con junto de datos simulado a partir de galaxias y nebulosas, sobre el cual se aplicaron distintos algoritmos de reconstrucción evaluados mediante métricas objetivas, que fueron utilizados como insumo para en trenar redes neuronales convolucionales diseñadas con y sin incorporación de contexto. Los modelos obtenidos demostraron capacidad para predecir métricas de calidad y ordenar reconstrucciones de manera consistente con las evaluaciones de referencia. Además, se realizó un análisis comparativo de algoritmos de reconstrucción considerando calidad de imagen, costo computacional y condiciones de observación. La principal ventaja de la solución propuesta es que aporta objetividad y eficiencia al análisis, reduciendo la carga de trabajo manual y permitiendo concentrar la atención en las re construcciones de mayor fidelidad. Finalmente, se implementó una herramienta prototipo que integra simulación, reconstrucción, evaluación y modelado en un mismo entorno flexible, capaz de incorporar nuevos algoritmos y métricas, con potencial de aplicación en futuros contextos de investigación.
This work focuses on the problem of objectively evaluating the quality of astronomical images reconstructed from interferometric observations, a task that has traditionally relied on expert visual inspection and has often been conducted without reproducible methodologies. To mitigate this li mitation, we propose an automatic system based on Deep Learning techniques, designed to compare multiple reconstructions from the same observation and to establish a quality ranking without the need for a reference image. To this end, a simulated dataset comprising galaxies and nebulae was generated, on which several reconstruction algorithms were applied and assessed using objective quality metrics; the resulting evaluations were subsequently employed to train convolutional neural networks, imple mented both with and without contextual information. The trained models demonstrated the capability to predict quality metrics and to order reconstructions consistently with reference-based evaluations. Furthermore, a comparative analysis of reconstruction algorithms was performed, considering image quality, computational cost, and observational conditions, thereby highlighting trade-offs between fi delity and efficiency. The principal advantage of the proposed approach is that it introduces objectivity and efficiency into the evaluation process, reducing manual effort and enabling experts to focus on the reconstructions of highest fidelity. Finally, a prototype tool was developed that integrates simulation, reconstruction, evaluation, and modeling into a unified and extensible framework, facilitating the in corporation of additional algorithms and metrics and offering potential applications in future research settings.
This work focuses on the problem of objectively evaluating the quality of astronomical images reconstructed from interferometric observations, a task that has traditionally relied on expert visual inspection and has often been conducted without reproducible methodologies. To mitigate this li mitation, we propose an automatic system based on Deep Learning techniques, designed to compare multiple reconstructions from the same observation and to establish a quality ranking without the need for a reference image. To this end, a simulated dataset comprising galaxies and nebulae was generated, on which several reconstruction algorithms were applied and assessed using objective quality metrics; the resulting evaluations were subsequently employed to train convolutional neural networks, imple mented both with and without contextual information. The trained models demonstrated the capability to predict quality metrics and to order reconstructions consistently with reference-based evaluations. Furthermore, a comparative analysis of reconstruction algorithms was performed, considering image quality, computational cost, and observational conditions, thereby highlighting trade-offs between fi delity and efficiency. The principal advantage of the proposed approach is that it introduces objectivity and efficiency into the evaluation process, reducing manual effort and enabling experts to focus on the reconstructions of highest fidelity. Finally, a prototype tool was developed that integrates simulation, reconstruction, evaluation, and modeling into a unified and extensible framework, facilitating the in corporation of additional algorithms and metrics and offering potential applications in future research settings.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Informático/a.
Keywords
Astronomía Imágenes, Astronomía Procesamiento de datos, Inteligencia artificial