Rapid Hyperspectral Classification of Camouflaged Objects in Complex Environments

Eunyeob Kim, Sungho Kim

Keywords: Hyperspectral Image Classification, Camouflaged Object Detection, CNN-Transformer Hybrid, Efficient Inference, Hyperspectral Dataset

TL;DR: We propose a fast hyperspectral object classification framework that enables multi-pixel prediction, achieving 7.44× faster scene-level inference while accurately detecting camouflaged objects in complex environments.

Abstract:

\begin{abstract}
Hyperspectral object classification has predominantly focused on satellite-based top-down imagery for large-scale land-cover classification. However, hyperspectral object classification in ground-based natural scenes remains largely underexplored due to the lack of publicly available datasets and the complexity of real-world environments.
In this work, we construct a hyperspectral dataset simulating defense scenarios where hazardous targets may be concealed within complex environments such as vegetation, trees, and camouflage structures. The constructed dataset consists of hyperspectral imagery acquired in the Visible–VNIR spectral range (479–900 nm), enabling scene-level analysis of concealed objects in natural environments.
Conventional hyperspectral classification models typically process hyperspectral cubes in a patch-wise manner. While effective for local feature extraction, this approach requires repeated patch inference to cover the entire scene, resulting in long inference times for scene-level analysis. To address this limitation, we propose a time-efficient inference framework
for fast scene-level hyperspectral object classification.
The proposed method enables the model to predict multiple pixels simultaneously
from a single hyperspectral patch, significantly reducing the number of required
inference operations. Experimental results demonstrate that the proposed method reduces inference time by 86.56\% (7.44$\times$ speedup) while reliably classifying concealed hazardous objects under complex and adverse environmental conditions.
\end{abstract}

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 14