A lightweight background suppression module for animal detection using remote sensing imagery

Deep learning-based animal detection using remote sensing imagery has achieved increasing attention in recent years, as it enables automated and large-scale monitoring in ecological applications. However, detecting animals in complex natural environments remains challenging due to visual similarity between animals and their surroundings. To alleviate this, we proposed a lightweight background suppression module (BSM) that enhances foreground features by suppressing background noise in the frequency domain. Specifically, the BSM introduces a learnable filtering mechanism to suppress background noise by adjusting adaptively the spectral magnitude of feature maps with channel-wise weights. A residuallike fusion strategy is then employed to combine adaptively the filtered and original features to enhance essential spatial details while reducing background noise. The BSM can be embedded into existing object detection frameworks. Experiments on an elephant detection dataset demonstrate that the BSM can increase both localization and counting accuracy with low computational cost compared to the baseline. Moreover, the BSM-enhanced model can maintain great performance under different animal densities. This research contributes a simple yet effective BSM for further detecting animals accurately under challenging remote sensing scenarios.