水资源是人类赖以生存的基础，精准的水面漂浮物检测是环保无人船进行河道清洁作业的首要前提。而目前的水面漂浮物检测任务仍面临诸多挑战。本文以提高水域分割和水面漂浮物检测的准确性和高效性为目标，旨在设计一种精准高效的水面漂浮物检测算法。论文研究的主要内容如下:
(1)针对复杂天气下无人船水域图像特征信息利用率低，轮边缘分割精度低的问题，设计改进PSPNet水域分割算法。首先，提出了一种交叉变换的通道特征金字塔模块 CT-CFP，该模块能够使不同层之间的特征信息交叉融合，提高原始特征信息的利用率;其次，设计了一种并行语义分割网络CFP-PSPNet，该网络通过PPM和CT-CFP双金字塔对图像信息进行提取，解决了细节信息和边缘信息丢失的问题;最后，将引入ECA 通道注意力机制的Mobilenetv2作为特征提取网络，在不影响分割精度的前提下，减少了网络的参数量和计算量，实现了网络的轻量化设计。

(2)针对目前目标检测算法对水面小目标的适应性差，特征识别能力低的问题，设计改进 YOLOv8 水面漂浮物检测算法。首先，设计了一种新的C2f-foat 模块，该模块通过将 bottleneck层输出的特征信息融合拼接，提升了特征信息利用率:其次，设计了一种 GS-EVC模块，该模块通过引入GSConv和shuMc操作提高了对水面漂浮物原始特征信息的利用率，加强了远程特征信息之间的依赖关系，提升了特征识别能力:最后将骨干网络中的普通卷积替换为ODConv，其中的加权注意机制能够适应复杂目标的特征提取，从而进一步提升网络检测精度。
(3)使用公开数据集对本文设计算法进行实验验证。结果表明:改进后的水域分割算法检测速度提升 8IFPS，平均交并比和平均精度分别达到97.71%和98.75%。改进后的水面漂浮物检测算法平均检测精度mAR。和mAR。9;分别提升了 4.3%和6.1%。实验证明本文设计的算法模型具有较好的鲁棒性和适用性，对无人船水面漂浮物检测的研究具有参考意义和应用价值。

Water resources are the basis for human survival, and accurate detection offloating objectson the water surface is the first prerequisite for environmental protection unmanned ships to carryout river cleaning operations. And the current task of detecting floating objects on the surface ofthe water still faces many challenges. In order to improve the accuracy and efficiency of watersegmentation and detection of floating objects on water surface, this paper aims to design anaccurate and efficient detection algorithm offloating objects on water surface. The main contentsofthis paper are as follows:
(1) Aiming at the problems of low utilization rate offeature information ofunmanned vesselwater image and low accuracy of contour edge segmentation under complex weather, theimproved PSPNet water segmentation algorithm is designed. Firstly, a cross-transform channelfeature pyramid module CT-CFP is proposed, which can improve the utilization rate of theoriginal feature information by cross-fusing the feature information between different layers.Secondly, a parallel semantic segmentation network CFP-PSPNet is designed, which extractsimage information through PPM and CT-CFP double pyramid to solve the problem of loss ofdetail information and edge information. Finally, Mobilenetv2, which introduces the attentionmechanism of ECA channel, is used as the feature extraction network, On the premise of notaffecting the segmentation accuracy, the number of parameters and calculation amount of thenetwork are reduced, and the lightweight design ofthe network is realized.

(2) Aiming at the problems of poor adaptability to small targets on the water surface andlow feature recognition ability in the current target detection algorithms, we design to improvethe YOLOv8 water surface floating object detection algorithm. Firstly, a new C2ffloat moduleis designed, which improves the feature information utilization by fusing and splicing the featureinformation output from the bottleneck layer. Secondly, a GS-EVC module is designed. By relationship between remote feature information, and improves the feafure recognition abilityFinally, the ordinary convolution in the backbone network is replaced with ODCony, in whichthe weighted attention mechanism can adapt to the feature extraction of complex targets, so asto further improve the detection accuracy of the network.

(3) The algorithm designed in this paper is experimentally validated using publicly availabldatasets. The results show that the improved water segmentation algorithm increases thedetection speed by 8lFps, and the average intersection ratio and average accuracy reach 97.71%and 98.75%, respectively, The average detection accuracy and average precision ofthe improvedfloating objects detection algorithm are inereased by 4,3% and 6.1%, respectively, Theexperiment proves that the algorithm model designed in this paper is robust and applicable.which is of reference significance and application value for the research of unmanned shipsurface floating object detection.