Skip to main content
Back

RPCASSM Spots Tiny Drones in Infrared by Rethinking How AI Scans Images

RPCASSM rewires Mamba-style state space models around classical RPCA math to sharply separate infrared backgrounds from tiny, dim targets

Paper, Full Access
  • New architecture: RPCASSM redesigns Mamba-style state space models using Robust PCA math to separate infrared backgrounds from tiny targets.
  • Two dedicated modules: BSSM scans for structured backgrounds; TSSM uses a deformable scan path targeting sparse, bright pixel regions.
  • Core problem solved: Generic vision SSMs fail to accurately reconstruct edges of targets occupying only a few pixels in infrared imagery.
  • Benchmark wins: RPCASSM outperforms CNN, transformer, and prior RPCA-SSM baselines on NUDT-SIRST and IRSTD-1K datasets, especially at low false-alarm rates.
  • Use cases: Maritime and aerial surveillance, search-and-rescue, early-warning systems, UAV fire detection.
  • Open source: Code to be released at github.com/PepperCS/RPCASSM.

Spotting a small, dim object against a cluttered infrared background -- think a drone over a city skyline or a vessel on a choppy sea -- is one of the hardest perception problems in computer vision. The target might occupy just a handful of pixels, its edges blurring into thermal noise, while the background shifts constantly. A new paper proposes RPCASSM, a network that rethinks how state space models (SSMs) should be structured for this exact problem, and the results beat a field of strong baselines on two standard benchmarks.

Why standard vision SSMs fall short here

Infrared small target detection (IRSTD) is widely used in security surveillance, search and rescue, and forest fire prevention, benefiting from infrared imaging's ability to function effectively at night and under adverse weather conditions. The field has been chasing better backbone architectures for years.

Mamba -- a state space model with linear complexity for long-distance modeling -- has attracted attention for this task, but direct application is unsuitable since local features, which are critical to detecting small targets, cannot be fully exploited. The core issue is structural: mainstream vision SSMs are designed to scan images in fixed raster patterns (row by row, column by column), which works well for natural images but is a poor fit for infrared scenes where the target is a sparse bright speck and the background is a smoothly varying low-rank surface.

Existing approaches have yet to adequately address the effective utilization of edge information within global modeling frameworks, notably lacking a mechanism to explicitly integrate multi-scale edge priors into a selective state space model-based global modeling. The result is that even recent SSM-based detectors struggle to accurately reconstruct the boundary of a target that is only a few pixels wide.