基于注意力机制和空洞卷积的无人机图像目标检测

UAV IMAGE OBJECT DETECTION BASED ON ATTENTION MECHANISM AND DILATED CONVOLUTION

  • 摘要: 针对现有无人机图像目标检测算法存在小目标检测精度低、多尺度目标漏检等问题,提出一种基于通道注意力机制和并行结构空洞卷积特征融合的无人机图像目标检测算法。该算法在ResNet50特征提取网络中引入SENet和PSDCFFN,从通道和感受野两个层面提高算法的特征表达能力,并使用ROI Align代替ROI Pooling,基于K-Means重新设计RPN(Region Proposal Networks)锚框尺寸,减小目标回归过程的坐标偏差。实验表明,该算法能够提升无人机图像目标检测精度,在RSOD-Dataset和无人机图像数据集上,mAP分别达到92.52%和98.07%。

     

    Abstract: The existing UAV image object detection algorithms suffer the problems of low detection accuracy on small objects and missed detection of multi-scale objects. A UAV image object detection algorithm based on channel attention mechanism and parallel structure dilated convolution feature fusion is proposed to overcome these problems. In order to improve the feature expression ability of the algorithm from channel and receptive field, we redesigned the ResNet50 via adding the SENet and PSDCFFN to the backbone. We used ROI Align and redesigned the RPN anchor size via K-Means to reduce the coordinate deviation in the object regression process. The experimental results show that the proposed algorithm can improve the accuracy of object detection in UAV images. On RSOD-Dataset and UAV image data sets, the mAP of the proposed algorithm reaches 92.52% and 98.07% separately.

     

/

返回文章
返回