基于时频双域协同与语义增强的复杂水域漂浮物检测方法OA

To address the issues of significant scale variations,strong background interference,and high miss detection rates of small objects in floating debris detection under complex water environments,this paper proposes a Spatial-Frequency Enhancement Network(SFE-Net).First,a Spatial-Frequency A-daptive Convolution is proposed.By employing a spatial-frequency split strategy and an adaptive fusion scheme,it overcomes the limitation of the fixed receptive fields in traditional convolution,achieving the col-laborative extraction of spatial contextual and frequency texture features,thereby significantly enhancing the perception capability for multi-scale objects.Second,a Wavelet Frequency Decoupling Module is de-signed,which decouples features into high-frequency details and low-frequency contours using Haar wavelet transform,and employs dual-path attention mechanisms to separately enhance edge textures and global semantic representations,effectively suppressing complex background noise such as water ripples and reflections.Finally,an Inner-IoU bounding box regression loss function is introduced.By incorpora-ting auxiliary bounding boxes and controlling scaling factors,it optimizes the internal similarity measure-ment of samples in feature space,addressing the problem of insufficient IoU sensitivity in small object lo-calization,and improving the convergence speed and localization accuracy of bounding box regression.Ex-perimental results on the FloW+dataset demonstrate that the proposed method achieves a mean Average Precision of 91.4%,representing an improvement of 8.7 percentage points over the original YOLOv8n.Meanwhile,the model parameters are reduced by 36.7%,and the inference speed reaches 214 frames per second,significantly enhancing detection performance while ensuring real-time capability.