Abstract
Infrared small target detection (IRSTD) faces significant challenges due to the minute size of targets and the high variability of imaging conditions across different sensor platforms and spectral bands. Traditional "visual-only" deep learning models often encounter parameter conflicts when trained on such diverse "omni-domain" data. This paper introduces AuxDet, a novel framework that incorporates auxiliary metadata—such as sensor platform, spectral band, and resolution—to dynamically guide feature learning. By leveraging metadata to resolve domain ambiguities, AuxDet achieves state-of-the-art performance on the WideIRSTD-Full benchmark, outperforming existing methods by over 30% in detection accuracy.
Introduction
The detection of small targets in infrared imagery is a critical task for surveillance and remote sensing. However, these targets often occupy only a few pixels, lack distinct textures, and are frequently obscured by complex background clutter. A major limitation of current state-of-the-art detectors is their reliance solely on pixel data. In real-world "omni-domain" scenarios, where images are captured across varying platforms (land, air, space) and spectral bands (LWIR, SWIR, NIR), a single visual-only model struggles to generalize. Features optimized for one domain may conflict with those of another, leading to poor overall performance.
The AuxDet framework addresses this by treating metadata as a high-level abstraction of the imaging environment. Instead of training separate models for each domain, AuxDet uses metadata to adapt a single model's "vision" to the specific physical context of each image, enabling cross-domain knowledge sharing.
Methodology
AuxDet introduces three primary components to integrate metadata into the detection pipeline:
1. Heterogeneous Metadata Encoding
To make metadata usable by a neural network, AuxDet converts categorical attributes (e.g., "Sensor: Space") and continuous values (e.g., "Resolution: 1024x1024") into high-dimensional latent vectors. This encoding process allows the model to ingest the physical context of the capture alongside the raw image data.
2. Multi-Modal Dynamic Modulation (M2DM)
M2DM serves as the core adaptive mechanism. It uses the encoded metadata to perform instance-specific modulation of the network's features. Rather than applying static filters, M2DM dynamically adjusts channel-wise and spatial features based on the metadata. This ensures that the network prioritizes features relevant to the specific imaging domain, such as looking for point-like targets in space-based SWIR imagery versus spot-like targets in land-based LWIR.
3. Lightweight Edge Enhancement Module (LEEM)
Small targets are easily lost in the deeper layers of standard convolutional networks. LEEM is a 1D convolutional block initialized with Laplacian-of-Gaussian (LoG) priors. By decomposing 2D filters into 1D horizontal and vertical chains, LEEM provides curvature sensitivity and amplifies high-frequency edges with minimal computational overhead, preventing target "wash-out."
Experimental Results
The authors evaluated AuxDet on the WideIRSTD-Full benchmark, a comprehensive dataset comprising 11,000 images from multiple public sources and new data. The dataset spans diverse platforms and spectral bands, providing a rigorous test for omni-domain detection.
AuxDet achieved a state-of-the-art 77.9% AP50 and 87.2% Recall. Notably, the inclusion of metadata-driven guidance resulted in a performance boost of over 30% compared to visual-only baselines. Furthermore, the framework demonstrated high flexibility; when integrated into various backbones like ResNet, Swin Transformer, or PVTv2, it consistently improved detection accuracy across the board.
Conclusion
AuxDet demonstrates that auxiliary metadata is a powerful, yet previously overlooked, resource for infrared small target detection. By moving from passive visual processing to context-aware active modulation, the framework successfully unifies multiple imaging domains into a single, high-performing model. Future research will focus on handling scenarios with missing metadata and extending the approach to temporal data in video sequences.
Summary updated