首页|基于注意力机制和多空间金字塔池化的实时目标检测算法

基于注意力机制和多空间金字塔池化的实时目标检测算法

扫码查看
原文链接
  • 万方数据
  • 维普
YOLOv4计算复杂度高，空间金字塔池化模块仅一次增强特征融合网络的深层区域特征图的表征能力、检测头网络的特征图难以突出重要通道特征;针对以上问题，提出一种基于注意力机制和多空间金字塔池化的实时目标检测算法;该算法采用多空间金字塔池化，提取局部特征和全局特征，融合多重感受野，加强特征融合网络的浅、中、深层特征图的表征能力;引入压缩激励通道注意力机制，建模通道间的相关性，自适应调整特征图各个通道的权重，从而使网络更加关注重要特征;特征融合和检测头网络中使用深度可分离卷积，减少了网络参数量;实验结果表明，所提算法的均值平均精度均高于其他7种主流对比算法;与YOLOv4相比，参数量、模型大小分别减少了 27。85 M和106。25 MB，所提算法在降低复杂度的同时，提高了检测准确度，且该算法的检测速率达到33。70帧/秒，满足实时性要求。
Real-Time Object Detection Algorithm Based on Attention Mechanism and Multi-spatial Pyramid Pooling
Aimed at the disadvantages of an enhancement to the representation of deep feature map in the enhanced feature fusion network for the spatial pyramid pooling module,higher computational complexity,and difficulty in highlighting important channel features for the feature map of the detection head network in YOLOv4 algorithm,Based on this problems,a real-time object detection algorithm based on attention mechanism and multi-spatial pyramid pooling is proposed.This algorithm adopts multi-spatial pyramid pooling,extracts the local and global features,fuses multiple receptive fields,and strengths the characterization ability of the shal-low,middle and deep feature maps for the feature fusion network.The squeeze-and-excitation channel attention mechanism is intro-duced to model the relativities between channels,the weight of each channel is adaptively recalibrated to make the network pay more attention to important features.Moreover,the deep separable convolution is used to reduce the parameters of the feature fusion and detection head networks.The experimental results show that the mean average precision(mPA)of the proposed algorithm is higher than that of other 7 mainstream comparison algorithms,compared with YOLOv4,the parameters and model size are reduced by 27.85 M and 106.25 MB,respectively.The proposed algorithm not only improves the detection accuracy,but also reduces the computation-al complexity compared to the baseline algorithm,and the average speed of the algorithm reaches by 33.70 FPS,which meets the re-al-time requirement.

YOLOv4channel attentionspatial pyramid poolingreceptive fielddepthwise separable convolutionreal-time

王国刚、李泽欣、董志豪

展开 >

山西大学 物理电子工程学院,太原 030006

YOLOv4 通道注意力 空间金字塔池化 感受野 深度可分离卷积 实时性

国家自然科学基金山西省自然科学基金山西省自然科学基金山西省高校科技创新计划山西省高校科技创新计划

11804209201901D111031201901D2111732019L00642020L0051

2024

10.16526/j.cnki.11-4762/tp.2024.02.009

计算机测量与控制
中国计算机自动测量与控制技术协会

计算机测量与控制

CSTPCD
影响因子：0.546
ISSN：1671-4598
年,卷(期)：2024.32(2)
  • 3