A new method for small-target water-floating garbage detection was proposed to address the issue of missed and false detections due to the limited information of small-target water-floating garbage that is prone to losing detailed features in feature extraction. After analyzing the limitations of small-target feature extraction, an edge-enhanced feature extraction network which synergistically utilized the spatial features and high-frequency features was proposed to effectively enhance the complex edges, details, and high-frequency information of small targets. A triple-branch cross-scale feature adaptive fusion module was designed to enhance the network’s ability to represent multi-dimensional small-target features by analyzing small-target features through local detail perception, global context modeling, and the large receptive field. An adaptive sparse attention-based intra-scale feature interaction module was constructed, which dynamically adjusted the interactive features by leveraging sparsity to enhance the discriminability between small targets and backgrounds. Experimental results show that, compared with the baseline model RT-DETR, the proposed method achieves improvements of 4.93, 2.46, and 3.18 percentage points in mAP, mmAP, and recall rate (R) respectively on the water-floating garbage dataset from the Lanzhou section of the Yellow River, and achieves improvements of 3.39, 1.45, and 2.23 percentage points on the FloW-Img dataset. These results indicate that the proposed method effectively enhances the detection performance for small-target water-floating garbage, thereby facilitating the efficient monitoring and management of water-floating garbage.