This paper introduces a lightweight model designed for underwater vehicles, utilizing Region Scaling (RS) loss and self-attention mechanisms to improve small-object detection in complex environments.

Replaces standard loss functions to better handle small or multi-scale objects.

Integrated into the neck or head of the network to capture global context without the heavy computational cost of standard transformers.

Published in the IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , this paper focuses on remote sensing and landslide detection using a modified YOLOv5/v10-style architecture. Full Text Access: Available via IEEE Xplore.

In these "Lightweight" (LS) models, the following components are typically highlighted in the full papers:

Reduces parameters and FLOPs while maintaining feature extraction quality.