Speaker
Description
With the arrival of Euclid/LSST and other large-scale surveys we address the automatic detection and segmentation of galactic features from deep sky images. The training of machine learning and deep learning systems requires manual annotations, which tend to present a high variability between annotators. For complex astrophysical features such as low surface brightness collision debris, even expert annotators do not perfectly agree on the features' exact shape and/or nature.
To avoid any ambiguity in the learning process, we propose to exploit the confidence information that is carried by the inter-annotator variability. We define a consensus associated with a new dedicated loss function, allowing us to exploit the inter-annotator variability. This loss mitigates learning based on a pixel-basis confidence measure. Annotators may be weighted within this consensus with respect to their expertise.
We experiment with various consensus formulas and galactic features to assess the effectiveness of this strategy in improving the learning of a deep neural network. A first series of experiments investigates the role of annotator composition and weighting within the consensus. We compare configurations ranging from the full annotator pool, to expert-weighted consensus, expert-only subsets, single-annotator baselines, and non-expert-only groups. This allows us to isolate the respective contributions of annotator diversity, expertise, and consensus formulation to the learning dynamics.
A second series of experiments examines the influence of network-level and data-level factors. We study how the strength and scale of input image normalisation affect feature representation, distinguishing between two complementary approaches: intensity scaling through arcsinh and sigmoid transforms, which compress the dynamic range of pixel values, and multi-scale decomposition through wavelet transforms, which expose structural information across spatial frequencies. We additionally evaluate data augmentation strategies, to identify those most beneficial for the segmentation of faint and diffuse structures.
Across both axes of experimentation, we obtain improved convergence and accuracy for the segmentation of various structures including low brightness structure.