Lightweight underwater debris detection model based on improved RT-DETR

As underwater ecosystems face escalating threats from increasing anthropogenic debris, autonomous monitoring and removal have become critical. Here we present LCSA-DETR, a lightweight object detection model optimized for deployment on resource-constrained autonomous underwater vehicles (AUVs). Built upon RT-DETR, LCSA-DETR introduces four core modifications to improve efficiency and detection performance: (i) replacement of the ResNet-18 backbone with StarNet to reduce computational complexity; (ii) a lightweight cross-stage aggregation encoder (LC-Encoder) with skip connections and feature alignment to enhance parameter efficiency; (iii) an adaptive kernel fusion block (AKFB) for improved multi-scale feature representation; and (iv) a bidirectional feature pyramid network (BiFPN) with dynamic weighting to enable effective feature fusion. In addition to these architectural enhancements, we apply a layer-adaptive magnitude-based pruning (LAMP) method to further compress the model, resulting in a pruned version named LCSA-DETR-P. Evaluated on the Trash-ICRA19 dataset without pretrained weights, LCSA-DETR-P achieves 80.4% AP and 98.3% AP50, delivering accuracy comparable to the baseline RT-DETR-R18 while significantly reducing computational demands. Specifically, it reduces the number of parameters, FLOPs, and model size to 31.7%, 23.0%, and 35.0%, respectively, and achieves a real-time inference speed of 105.1 frames per second, demonstrating its suitability for deployment in resource-constrained underwater scenarios.