高级检索

融合自适应加权模块的YOLOv4多尺度目标检测算法

YOLOv4 Multi-Scale Object Detection Algorithm Based on Adaptive Weighting Module

  • 摘要: 针对残差块中特征图直接相加无法反映特征间相关性以及特征融合阶段存在信息损失的问题, 提出一种融合自适应加权模块的目标检测算法. 该算法对YOLOv4的主干网络和颈部网络进行改进, 设计一种自适应加权模块, 对输入的两个特征图进行通道压缩并提取特征信息, 接着从提取的信息中获取空间权重, 最后将空间权重赋予输入特征图来建立它们之间的相关性. 在特征融合阶段增加跳跃连接, 补充特征融合过程中的信息损失. 实验结果表明, 所提算法能有效提高目标的检测精度, 在PASCAL VOC数据集与KITTI数据集上的mAP值较YOLOv4算法分别提高0.90%与0.89%.

     

    Abstract: The direct addition of feature maps in residual block cannot reflect the correlation between features and there is information loss in feature fusion stage. Therefore, an object detection algorithm based on adaptive weighting module is proposed. The algorithm improved the backbone network and neck network of YOLOv4, and an adaptive weighting module is designed to compress the channel of two input feature maps and extract feature information, the spatial weights are then obtained from extracted information, and finally the spatial weights are assigned to the input feature maps to establish correlation between them. In the feature fusion stage, skip connections are added to supplement information loss in feature fusion process. Experimental results show that the proposed algorithm can effectively improve detection accuracy, the mAP values of the proposed algorithm on PASCAL VOC and KITTI datasets are 0.90% and 0.89% higher than that of YOLOv4, respectively.

     

/

返回文章
返回