基于改进YOLOv5s的水面垃圾检测方法

Water surface garbage detection based on improved YOLOv5s

针对传统水面垃圾检测方法在水面环境复杂的情况下难以兼顾检测精度和检测速度以及检测效率低的问题,提出一种改进YOLOv5s的水面垃圾检测方法。首先使用GhostBottleneck模块替换YOLOv5s模型结构中的C3和部分CBS模块,使YOLOv5s模型更加轻量化,提升检测效率;然后在骨干网络的末端加入ECA注意力机制,让模型对水面垃圾的特征提取有更多关注;再使用Alpha-CIOU损失函数代替YOLOv5原本的GIOU损失函数,使回归更加准确,以获取质量更好的Anchor Box。经过上述改进获得YOLOv5s-GEA模型,模型规模减少了11.2%,每秒帧率FPS为41.6,在测试集上的mAP值达到90.3%,相比原YOLOv5s模型提高了3.9%。相比传统的水面垃圾检测方法,提出的YOLOv5s-GEA模型可以有效兼顾检测精度和检测速度,能够准确迅速地检测出水面垃圾的种类和位置,并且具有较小的模型体积,易于部署到水面清理机器移动端。

Addressing the challenges faced by conventional water surface garbage detection techniques in reconciling detection accuracy and speed,as well as the low detection efficiency in intricate aquatic environments,we present an enhanced water surface garbage detection approach founded on the YOLOv5s framework.Initially,the GhostBottleneck module substitutes the C3 and a portion of CBS modules within the YOLOv5s model structure,rendering the YOLOv5s model more compact while augmenting detection efficiency.Subsequently,the incorporation of the ECA attention mechanism at the conclusion of the backbone network enables the model to focus more intently on water surface garbage feature extraction.Lastly,the Alpha-CIOU loss function supplants the YOLOv5s original GIOU loss function,resulting in more precise regression and superior quality Anchor Boxes.The resulting YOLOv5s-GEA model demonstrates a reduction in model size by 11.2%,a frame rate of 41.6 FPS,and a mAP value of 90.3%on the test set,constituting a 3.9%enhancement in comparison to the unmodified YOLOv5s model.This evidence indicates that,relative to traditional water surface garbage detection methods,the proposed YOLOv5s-GEA model proficiently balances detection accuracy and speed,precisely and expeditiously determining the types and locations of water surface garbage.Moreover,it has a smaller model size,making it easy to deploy on mobile devices for water surface cleaning machines.