| 摘要: |
| 为解决航天遥感图像分辨率和目标尺度变化大的挑战，提出了一种基于多分辨率图像的目标检测算法。改进了自适应特征金字塔和轻量级的分类预测模块，通过使用注意力机制，从不同层次的特征图中提取语义信息。引入了一种预测目标尺度的方法，以分析图像中目标的分布和尺度信息。将算法在DOTA（Dataset for Object deTection in Aerial Images）数据集上进行了实验验证，在U-Net（一种基于卷积神经网络的语义分割算法）和ResNet-34（一种深度残差网络）两种不同的主干网络设置下，召回率和检测速度均超过了RPN（Region Proposal Network，区域提议网络）算法。提出的多分辨率图像目标检测算法能有效地提高检测精度，降低计算复杂度。 |
| 关键词: 深度神经网络 目标检测 特征金字塔 多尺度目标 多分辨率图像 |
| DOI： |
|
| 基金项目: |
|
| A Multi-Resolution Image Object Detection Algorithm |
| DU Linlin,ZHANG Yu,TANG Yu,MA Jingjing |
| (Systems Engineering Institute,Academy of Military Science, PLA, Beijing 100010, China) |
| Abstract: |
| To address the challenges of large image resolution and target scale variations in remote sensing images, this work proposes a target detection algorithm based on multi-resolution images. The adaptive feature pyramid and lightweight classification prediction module are improved. By using attention mechanisms, we extract semantic information from feature maps at different levels，introduce a method for predicting target scale to analyze the distribution and scale information of targets in the images.The algorithm is experimentally validated on the DOTA dataset. With two different backbone network settings, U-Net and ResNet-34, the recall rates and the detection speeds both surpassing the RPN algorithm. The proposed multi-resolution image target detection algorithm effectively improves detection accuracy while reducing computational complexity. |
| Key words: Deep neural networks Object detection Feature pyramids Multi-scale object Multi-resolution image |