Abstract: To address the issue of low accuracy in detecting small floating objects on a reservoir, we proposed a YOLOX-based detection framework for water surface floating object recognition.The proposed detector introduces a novel dark2 module, which was embedded into the backbone as a plug-and-play module, to develop the branch structure and enhance feature extraction and representation for given images.Furthermore, we designed a modified feature aggregation module(ZL-FPN) to facilitate the fusion and interaction of multi-scale features, and the detection accuracy of small floating objects on a reservoir was improved.The results demonstrated that the proposed model obtained 29.93% and 12.11% performance gains compared with YOLOv4 and the original YOLOX.The research findings can provide effective technical support for improving the level of intelligent management of reservoirs.