LRS-RT-DETR: A Long-Range Floating Garbage Detector with Multi-Scale Feature Fusion

Abstract

Real-time detection of floating water surface garbage is of great significance for water environment management. However, when the garbage target is more than 30 meters away from the camera, the target occupies only a small number of pixels in the image, and the complex water surface background also introduces detection interference; thus, general object detection methods struggle to achieve satisfactory performance. In this paper, we propose a long-range small object detection Transformer model named LRS-RT-DETR, which is improved based on the RT-DETR-R18 baseline, to adapt to the scenario of long-distance small garbage detection on water surfaces. First, we design a high-resolution feature fusion pyramid network called High-Resolution Feature Fusion Pyramid Network (HRFF-FPN). By employing the lossless Space-to-Depth (SPD) reorganization operation, we introduce the high-resolution P2 feature into the P3 detection node, significantly enhancing the model’s spatial perception capability for extremely small distant targets without adding extra detection heads. After receiving the fused features from the P2 layer, we propose a multi-scale kernel branch module (MSKB), which combines skip connections with a multi-scale parallel receptive field branch to achieve multi-level refined modeling of the fused features. Experimental results on the targeted long-distance water surface garbage dataset Far-water-surface Garbage Dataset (FWSGD) show that LRS-RT-DETR achieves 91.5% mAP@0.5, an improvement of 2.7 percentage points over the baseline RT-DETR-R18, and 43.9% mAP@0.5-0.95, an improvement of 1.5 percentage points. Meanwhile, the model parameter count increases by only 0.6M, with controllable computational overhead, demonstrating good potential for real-time deployment on edge devices.