Resolve dense-target infrared detection limitations

Establish methods that overcome the local spatial saliency, density-aware semantic guidance, and dense-target structural limitations that cause target adhesion, incomplete separation, ambiguous spatial structures, mutual feature interference, inaccurate localization, and unstable detection in densely distributed infrared small-target scenes.

Background

The paper identifies dense infrared small-target detection as a distinct unresolved setting. Existing methods based on local spatial saliency have difficulty distinguishing neighboring targets from their backgrounds, while current deep neural networks do not explicitly encode target density during feature extraction. In addition, models developed for sparse targets do not represent the spatial hierarchy of densely distributed targets, leading to ambiguous structures and feature interference.

These limitations affect target separation, localization accuracy, and detection stability in scenes containing clustered or otherwise densely distributed infrared targets. The authors therefore present the three deficiencies as current unsolved challenges requiring dedicated solutions.