Objective In complex inland river environments most floating garbage consists of small targets that are easily affected by reflections from the water surface and riverbanks. This results in blurred and obstructed target shapes which poses significant challenges for target detection. To address this issue this paper proposed an improved YOLOv7-based algorithm for detecting floating garbage in rivers. Methods Firstly to address the morphological variations of floating garbage in river environments caused by environmental factors the feature extraction capabilities for small targets were enhanced through the refinement of the SPPCSPC module. Secondly the centralized feature pyramid was added which was weighted and fused with the feature pyramid through ROI region of interest to facilitate the detection of targets at different scales. Finally given that traditional Intersection over Union IoU is highly sensitive to positional deviations of small targets which reduces detection performance IoU was replaced with Wasserstein Distance as the evaluation metric. A loss function based on Normalized Wasserstein Distance NWD was implemented to improve detection accuracy. Results The experimental results showed that the accuracy of the improved YOLOv7 algorithm model increased by 3. 1% reaching 89. 7%. At IoU = 0. 5 the average mean precision increased by 6% reaching 87. 8% and for IoU ranging from 0. 5 to 0. 95 the average mean precision increased by 4. 6% reaching 43. 4%. The detection results of this improved model outperform those of other classical detection models. Conclusion The experimental results indicate that the improved model significantly enhances detection accuracy providing valuable insights for practical applications.