基于轻量卷积和信息增强的目标检测算法
Object Detection Algorithm Based on Lightweight Convolution and Information Enhancement
王惠杰 1李忠飞 1张云峰 1李明 1樊世君 1聂帅杰2
作者信息
- 1. 内蒙古电投能源股份有限公司北露天煤矿
- 2. 中国矿业大学信息与控制工程学院
- 折叠
摘要
为解决在矿井环境中目标检测算法模型体积大、计算复杂度高以及模型轻量化后精度低的问题,提出了一种专为矿井环境设计的目标检测算法——YOLO-AM.该算法采用轻量化网络MobileNetv2作为主干网络,并使用深度可分离卷积来替代颈部网络中的3×3卷积,显著降低模型的计算量和参数量.这一设计使得算法更适应矿井中有限的计算资源和对实时性的需求.接着在主干网络的输出位置引入坐标注意力机制,增强输出特征中的有效信息.同时提出了一种浅层特征增强模块,在特征融合网络融合该模块用于增强浅层特征的语义信息,从而提高模型的检测精度.在公共数据集PASCAL VOC上的试验结果表明,相比基准模型YOLOv4,YOLO-AM以降低7%检测精度的代价,减小了83%的参数量和86%的计算量,同时也提高了检测速度.
Abstract
In order to solve the problems of large volume,high computational complexity and low pre-cision of target detection algorithm model in mine environment,a target detection algorithm YOLO-AM spe-cially designed for mine environment is proposed.The algorithm adopts the lightweight MobileNetv2 as the backbone network and employs depth-wise separable convolution to replace the 3×3 convolution in the neck network,significantly reducing the model's computational and parameter complexity.This design enables the algorithm to better adapt to the limited computational resources and real-time requirements in mining en-vironments.Furthermore,a coordinate attention mechanism is introduced at the output position of the back-bone network to enhance the effective information in the output features.Simultaneously,a shallow feature enhancement module is proposed,which is integrated into the feature fusion network to augment the seman-tic information of shallow features,thereby improving the model's detection accuracy.Experimental results on the public dataset PASCAL VOC demonstrate that,compared to the baseline model YOLOv4,YOLO-AM achieves a 7%reduction in detection accuracy but significantly reduces parameter complexity by 83%and computational complexity by 86%.Additionally,it enhances detection speed.
关键词
矿井目标检测/YOLOv4/MobileNetv2/深度可分离卷积/注意力模块/特征融合/轻量化/主干网络Key words
mining object detection/YOLOv4/MobileNetv2/depth separable convolution/attention module/feature fusion/model lightweighting/backbone network引用本文复制引用
出版年
2024
NETL
NSTL
万方数据