| 摘 要: 针对水下图像模糊、纹理缺失及复杂背景干扰导致的检测难题，提出了一种基于双域频空融合与多尺度边缘增强的改进算法。该算法以YOLOv11为基准，首先设计了双域自适应频空融合模块，利用快速傅里叶变换捕获全局频域特征，增强了水下目标的高频纹理特征。其次，提出了多尺度RFA引导边缘上采样模块，提升了对微小及伪装目标的定位精度。最后，引入小波残差边缘下采样模块与几何感知混合损失函数，构建特征无损传输路径并优化边界框回归。在UTDAC数据集上实验表明，该模型在仅微增参数量的情况下显著提升了检测性能。 |
| 关键词: 深度学习 水下目标检测 YOLOv11 双域频空融合 边缘增强 |
|
中图分类号: TP391
文献标识码:
|
| 基金项目: 国家自然科学基金项目（62376109） |
|
| Lightweight Underwater Small Object Detection Based on Dual-Domain Frequency-Spatial Fusion and Multi-Scale Edge Enhancement |
|
LV Kangqi, WANG Weidong, QIN Bin
|
School of Computer, Jiangsu University of Science and Technology
|
| Abstract: To address the detection challenges caused by underwater image blurring, texture loss, and complex background interference, an improved algorithm based on dual-domain frequency-spatial fusion and multi-scale edge enhancement is proposed. Using YOLOv11 as the baseline, the algorithm first designs a Dual-domain Adaptive Frequency-Spatial fusion module, utilizing Fast Fourier Transform to capture global frequency domain features and enhance the high-frequency texture features of underwater objects. Second, a Multi-scale RFA-guided Edge Upsampling module is proposed to improve the localization precision of tiny and camouflaged objects. Finally, a Wavelet Residual Edge Downsampling module and a geometry-aware hybrid loss function are introduced to construct a lossless feature transmission path and optimize bounding box regression. Experiments on the UTDAC dataset demonstrate that the model significantly improves detection performance with only a marginal increase in parameter count. |
| Keywords: Deep learning Underwater object detection YOLOv11 Dual-domain frequency-spatial fusion Edge enhancement |
