Determine whether all model combinations have been explored for infrared small-target detection

Determine whether all combinations of decomposition models, minimization problems, loss functions, solvers, neural-network architectures, optimization procedures, loss functions, and training strategies have been investigated to identify suitable configurations for infrared small-target detection.

Background

The paper observes that model-based and data-based infrared small-target detection methods are constructed by combining multiple design choices. Low-rank and sparse decomposition methods vary in their decompositions, minimization problems, loss functions, and solvers, whereas deep neural-network methods vary in conceptual model, architecture, optimization, loss function, and training.

Because these choices produce a large configuration space, the authors question whether the field has examined all relevant combinations. They suggest that AI-assisted research tools could help address this uncertainty by searching for suitable combinations for the infrared small-target detection task.