基于MEFF-YOLO的海洋底栖生物目标检测算法

Marine Benthic Organism Detection Algorithm Based on MEFF-YOLO

摘要:

针对水下目标检测中图像模糊、生物群聚遮挡重叠问题,本研究提出一种基于YOLOv11-S改进的MEFF-YOLO算法。设计边缘特征融合主干网络(EFF-DarkNet),通过多尺度边缘特征生成模块(MEFG)提取浅层高分辨率边缘信息,并利用跨通道融合模块(EFFC)实现边缘特征与常规卷积特征的深度融合,提升目标边界表征能力。其次,提出多尺度部分聚集卷积模块(MPAC),通过分层级联卷积与残差连接,在减少冗余计算的同时保留多尺度原始信息。最后,提出Inner-MPDIoU损失函数,融合尺度自适应辅助边界框策略与最小点距优化方法,提升边界框定位精度。在DUO数据集上的实验表明,MEFF-YOLO以10.26M参数量实现72.1%的mAP0.5:0.95,推理速度达227.2 FPS,较YOLOv11-S精度提升3.3%,为复杂水下环境中的生物检测提供了高精度、高效率的解决方案。

Abstract:

Targeting the issues of image blurriness and biological cluster occlusion in underwater object detection, this study proposes an improved MEFF-YOLO algorithm based on YOLOv11-S. We designed an Edge Feature Fusion backbone network (EFF-DarkNet), which uses a Multi-scale Edge Feature Generation module (MEFG) to extract shallow high-resolution edge information, and leverages a Cross-Channel Fusion module (EFFC) to achieve deep integration of edge features and conventional convolutional features, enhancing the representation of object boundaries. Additionally, we introduced a Multi-scale Partial Aggregation Convolution module (MPAC), which retains multi-scale original information while reducing redundant computations through hierarchical concatenation and residual connections. Finally, we proposed the Inner-MPDIoU loss function, which combines scale-adaptive auxiliary bounding box strategy and minimum point distance optimization method to improve bounding box localization accuracy. Experiments on the DUO dataset show that MEFF-YOLO achieves an mAP0.5:0.95 of 72.1% with 10.26M parameters and a reasoning speed of 227.2 FPS, marking a 3.3% improvement in accuracy over YOLOv11-S. This provides a high-precision, high-efficiency solution for biological detection in complex underwater environments.