摘要
针对当前目标检测算法模型复杂度高、实时检测图片速度较慢的难题,提出一种基于跨通道融合与注意力机制的轻量化目标检测模型,便于移动场景搭载。利用MV1模块构建主干特征提取网络,CDC模块改造多层基础卷积,提出一种基于深度可分离卷积(depthwise separable convolution,DW)的轻量化模型。在网络的颈部,三通道扩展为四通道;同时设计一种跨通道路径,融合深浅层信息,进一步加强特征提取。最后融合一种高效通道注意力机制(efficient channel attention,ECA),使算法聚焦图像中的目标位置。仿真实验结果表明:该算法在Pascal VOC 07+12数据集上所有类别的平均识别准确率达到91.64%,与最新的YOLOv7相比,模型参数量降低了60%,检测速度达到55.54 FPS,证明了该算法的有效性。
In view of the high complexity of the current object detection algorithm model and the slow speed of real-time testing pictures,a lightweight target detection model based on cross channel fusion and attention mechanism is proposed,which is convenient for carrying mobile scenes.Firstly,MV1 module is used to build the backbone feature extraction network,CDC module is used to transform multi-layer basic convolution,and a lightweight model based on Depth Separable Convolution(DW)is proposed.Secondly,in the neck of the network,three channels are extended to four channels;Meanwhile,a cross channel path is designed to fuse deep and shallow information to further enhance feature extraction.Finally,an Efficient Channel Attention(ECA)is integrated to make the algorithm focus on the target position in the image.The simulation results show that the average recognition accuracy of all categories of the algorithm in Pascal VOC 07+12 dataset reaches 91.64%.Compared with the latest YOLOv7,the model parameters are reduced by 60%,and the detection speed reaches 55.54 FPS,the effectiveness of the proposed algorithm is proved.
作者
曹瑞颖
赵志诚
谢新林
刘宁
CAO Ruiying;ZHAO Zhicheng;XIE Xinlin;LIU Ning(School of Electronic Information Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,China;Shanxi Information Industry Technology Research Institute Co.,Ltd.,Taiyuan 030012,China)
出处
《火力与指挥控制》
CSCD
北大核心
2024年第3期142-150,共9页
Fire Control & Command Control
基金
山西省青年科学基金项目(201901D211304)
山西省科技重大专项(20191102009)。
