CGO: Multiband Astronomical Source Detection With Component-Graphs

Thanh Xuan Nguyen, Giovanni Chierchia, Laurent Najman, Aku Venhola, Caroline Haigh, Reynier Peletier, Michael H. F. Wilkinson, Hugues Talbot, Benjamin Perret

Introduction

Detecting faint stars and galaxies across multiple color bands is a fundamental challenge in astronomy. Traditional source detection tools, such as SExtractor or MTObject, typically process images one color band at a time or rely on "flattening" multiband data into a single grayscale image. These approaches often lose subtle information and struggle with the "partial ordering" inherent in multiband data—where a pixel's brightness may increase in one band while decreasing in another.

Illustration of partial ordering in multiband data, where intensity pairs (e.g., (2,1) vs (1,2)) cannot be strictly ordered, necessitating a graph structure.

To address these limitations, this paper introduces CGO (Component-Graph Object detection), a framework that utilizes Component-Graphs (CG) to analyze multiband images simultaneously. Unlike traditional Component Trees (CT), which are limited to total orderings in single-band data, CGs are Directed Acyclic Graphs (DAGs) that can naturally represent complex multiband relationships, improving detection sensitivity for the dimmest sources.

The CGO Framework

The CGO framework moves beyond simple tree structures to filter complex graphs for meaningful physical objects. The process involves several key innovations:

  1. Component-Graph Construction: A CG is built where nodes (sets of connected pixels) can have multiple parents across different bands, capturing the full spectral information.
  2. Generalized Significance Testing: The authors adapted statistical tests to a multiband context using a chi-square distribution. This allows the system to determine if a group of pixels represents a real signal by evaluating "node power" across all bands simultaneously.
  3. Node Differentiation and Deblending: To manage the complexity of DAGs, CGO uses a distance-based predicate to decide if incomparable nodes belong to the same physical object, effectively deblending overlapping sources.
  4. Partial Object Detection: A specialized module ensures that fragments of an object are correctly grouped together rather than being discarded as background noise.
The CGO framework pipeline: from multi-band input to knowledge space extraction via component-graph construction and filtering.

Experimental Evaluation

The researchers evaluated CGO against three distinct datasets:

  • Synthetic Data: 10,000 three-band images used to establish theoretical limits.
  • Simulated Optical Data: A simulation of the Fornax Deep Survey containing thousands of stars, galaxies, and background clusters.
  • Real Astronomical Data: Multiband images ($u, g, r, i$) from the Kilo-Degree Survey (KiDS).

Key Results

The evaluation demonstrated that CGO provides higher theoretical bounds for capturing the true shape of astronomical sources compared to Component Trees. In simulations, CGO consistently outperformed MTObject, achieving higher recall (detecting more faint objects) while maintaining superior precision (fewer false positives). Qualitative analysis of the KiDS dataset confirmed that CGO successfully segments complex, multi-colored galactic structures that traditional methods often fail to deblend correctly.

Conclusion and Future Work

By processing all bands simultaneously, CGO detects faint sources that might not meet significance thresholds in any single band but are statistically significant when their energy is summed across the spectrum. While the current implementation faces computational challenges due to the $O(n^2)$ complexity of graph construction—requiring image tiling—future work will focus on optimizing these algorithms and exploring additional attributes like shape and color to further refine source detection.

Summary updated

Abstract

Component-graphs provide powerful and complex structures for multi-band image processing. We propose a multiband astronomical source detection framework with the component-graphs relying on a new set of component attributes. We propose two modules to differentiate nodes belong to distinct objects and to detect partial object nodes. Experiments demonstrate an improved capacity at detecting faint objects on a multi-band astronomical dataset.
BibTeX
@inproceedings{nguyen2020cgo,
  author = {Thanh Xuan Nguyen and Giovanni Chierchia and Laurent Najman and Aku Venhola and Caroline Haigh and Reynier Peletier and Michael H. F. Wilkinson and Hugues Talbot and Benjamin Perret},
  title = {{CGO: Multiband Astronomical Source Detection With Component-Graphs}},
  booktitle = {ICIP},
  year = {2020},
  doi = {10.1109/icip40778.2020.9191276},
}

Loading related content...

Figure