基于改进YOLOv5s的水面漂浮物检测算法研究
Research on Detection Algorithm of Floating Objects on Water Surface Based on Improved YOLOv5s
项新建 1翁云龙 1谢建立 2郑永平 1吴善宝 1许宏辉 1杨斌2
作者信息
- 1. 浙江科技大学 自动化与电气工程学院,浙江 杭州 310023
- 2. 凯铭科技(杭州)有限公司,浙江 杭州 310023
- 折叠
摘要
水面图像具有水波扰动、光线反射、岸上倒影等复杂特征,导致现有目标检测方法不能很好地完成对水面漂浮物的识别.基于此,提出了一种基于改进YOLOv5s的水面漂浮物检测优化模型,通过增加目标检测层以提升模型检测多尺度目标的能力,在特征融合层引入无参注意力机制SimAM以提高模型对漂浮物特征的学习,并采用CARAFE上采样方式增强网络的感受野以提高对特征的重建能力,在YOLOv5s网络结构中融入卷积混合层以保持模型检测精度且减少参数量,从而提高模型运行速度.实例验证结果表明:改进模型检测效果良好,平均精度达 97.1%,较原YOLOv5s模型提高了 4.9 个百分点,能够有效改善水面漂浮物漏检、误检问题.
Abstract
The water surface image has complex features such as water wave disturbances,light reflection and shoreline reflections,which cause existing object detection algorithms to fail to recognize floating objects.Therefore,a water surface floating object detection algorithm based on an improved YOLOv5s was proposed.By modifying the neck structure of the network,adding object detection layers to improve the detection accuracy of the feature extraction network for multi-scale targets;introducing the parameter-free attention mechanism SimAM in the feature fusion layer to enhance the model's feature extraction ability,and using the CARAFE upsampling method to enhance the network's receptive field and improve the perception of detailed features;integrating the ConvMixer Layer into the YOLOv5s network structure,the model's running speed was improved while maintaining detection accuracy and reducing the model's parameter count.The experimental re-sults show that the improved model has good detection performance in real samples,with a mean average precision of 97.1%,which is 4.9%higher than the original YOLOv5s model,and can effectively improve the issue of missed detection and false detection of floating objects on the water surface.
关键词
漂浮物检测/YOLOv5s/多尺度特征检测/注意力机制/CARAFE/卷积混合层Key words
floating object detection/YOLOv5s/multi-scale feature detection/attention mechanism/CARAFE/ConvMixer Layer引用本文复制引用
基金项目
浙江省自然科学基金资助项目(LQ16F030002)
浙江省重点研发计划项目(202206)
杭州市科技计划发展项目(202203B21)
出版年
2024
国家科技期刊平台
NSTL
万方数据