Probing Deep into Temporal Profile Makes the Infrared Small Target Detector Much Better.

Area of Science:

• Computer Vision
• Signal Processing
• Artificial Intelligence

Background:

• Infrared small target (IRST) detection faces challenges with dim targets and interference.
• Current spatial-temporal methods are computationally redundant and unreliable in complex conditions.

Purpose of the Study:

• To explore essential information for IRST detection beyond spatial and short-term temporal domains.
• To develop a more efficient and robust IRST detection method by leveraging global temporal saliency.

Main Methods:

• Theoretical analysis revealed the superiority of temporal profile information for distinguishing targets.
• Developed the first prediction attribution tool to verify the importance of temporal data.
• Remodeled IRST detection as a 1D signal anomaly detection task.
• Proposed the Deep Temporal Probe Network (DeepPro) operating solely in the time dimension.

Main Results:

• DeepPro demonstrated significant superiority in distinguishing target signals using temporal saliency.
• Experimental validation confirmed DeepPro's effectiveness on widely-used benchmarks.
• Achieved state-of-the-art performance with extremely high efficiency, especially for dim targets and complex scenarios.

Conclusions:

• The temporal profile is a crucial domain for effective IRST detection.
• DeepPro offers a novel, efficient, and robust solution for IRST detection.
• This work promotes advancements in IRST detection through a new modeling approach and insights.