Abstract
Salient Object Detection (SOD) in the hyperspectral domain (HSOD) offers significant advantages over traditional RGB methods by utilizing material-specific spectral signatures. However, the field faces challenges due to data scarcity and the difficulty of processing high-dimensional data without losing spatial detail. We present HSOD-BIT-V2, the largest and most diverse HSOD dataset to date, and Hyper-HRNet, a high-resolution network featuring a novel attention-based reconstruction module and a ternary-aware decoder. Our approach sets new state-of-the-art benchmarks, particularly in detecting small objects and navigating complex backgrounds.
Introduction
Salient Object Detection (SOD) aims to identify the most visually prominent objects in a scene. While RGB-based SOD has matured, it struggles with challenges like low contrast, camouflage, and complex backgrounds. Hyperspectral Imaging (HSI) provides a solution by capturing hundreds of spectral bands, offering a "spectral fingerprint" for materials.
As shown in Figure 1, even when objects are visually similar to the background in pseudo-color representations, their spectral curves remain distinct, enabling more accurate detection where RGB methods fail. However, HSOD has been limited by small datasets and the "Hughes phenomenon," where high dimensionality can paradoxically degrade model performance.
Methodology: Hyper-HRNet
To address these issues, we propose Hyper-HRNet, designed to preserve high-resolution details while intelligently managing hyperspectral data.
Hyperspectral Attention Reconstruction (HAR)
Instead of traditional dimensionality reduction like PCA, which often discards salient signals, HAR uses a learnable Hybrid Perceptual Spectral Attention mechanism. It combines Transformer-based self-attention for long-range spectral dependencies with CNN-based adaptive attention for local spatial details, effectively compressing data while preserving saliency features.
Global Ternary Perception Decoder (GTPD)
To solve the "blurry boundary" problem common in upsampling, the GTPD utilizes a Ternary-Aware Weight (TAW) mechanism. By generating a "trimap" that explicitly classifies pixels as object, background, or uncertain boundary, the model learns to produce significantly sharper object contours.
The HSOD-BIT-V2 Dataset
The paper introduces HSOD-BIT-V2, which raises the bar for the field with 500 high-quality HSIs (1240 × 1680 resolution) across 200 spectral bands. It includes diverse natural backgrounds, such as snowfields and fallen leaves, and categorizes images by five difficulty factors: Complex Background (CB), Color Similarity (CS), High Dynamic Range (HDR), Small Object (SO), and Material Similarity (MS).
As illustrated in Figure 2, HSOD-BIT-V2 provides a much more rigorous testing ground than previous iterations, particularly in the realms of complex backgrounds and small object detection.
Key Results
Hyper-HRNet sets a new state-of-the-art across multiple benchmarks. On HSOD-BIT-V2, it achieves a Mean Absolute Error (MAE) of 0.028 and an AUC of 0.988, significantly outperforming both RGB-based models (e.g., U2Net) and previous HSI-based leaders. Furthermore, the Hyper-HRNet-Lite variant offers a computationally efficient alternative, maintaining competitive accuracy with only 7.24M parameters.
Conclusion
HSOD-BIT-V2 and Hyper-HRNet represent a significant step forward for hyperspectral saliency detection. By providing a diverse, high-resolution benchmark and an architecture that leverages spectral fingerprints while maintaining spatial precision, this work enables more robust object detection in complex real-world environments.