Abstract:

Predictive image detection based on deep learning has a wide range of applications in the fields of earth resource investigation, military reconnaissance, and environmental monitoring. More accurate and efficient target detection algorithms are the hotspots and difficulties of existing image detection research. An improved CenterNet algorithm is proposed to remote sensing image detection, which preprocesses the remote sensing image to adapt to the CenterNet network and improves the detection efficiency of the remote sensing image. The original network is improved, the standard convolution in the residual module is replaced by the depthwise separable convolutions, which effectively reduces the amount of network calculations and reduces redundancy. At the same time, an attention mechanism is added to suppress useless information and improves the accuracy of network detection. In view of the large observation area of ??the remote sensing image, the relatively small target, the large difference in target size and the uneven distribution, the false detection rate and the missed detection rate of the target are reduced. The experimental results show that the improved CenterNet algorithm has a significant improvement over the original CenterNet algorithm, which proves the robustness of the improved algorithm.

Key words: target detection, remote sensing images, depthwise separable convolutions, attention mechanism, Anchor-Free mechanism, CenterNet algorithm

摘要：

基于深度学习的遥感图像检测在地球资源调查、军事侦察、环境监测等领域有着广泛的应用，更精准、高效的目标检测算法是目前遥感图像检测研究的热点和难点。提出一种改进的CenterNet遥感图像检测算法，对遥感图像进行预处理，以适应CenterNet网络，提高网络对遥感图像的检测有效性；对原网络进行改进，将残差模块中的标准卷积替换成深度可分离卷积，有效降低网络计算量，减少冗余；同时加入注意力机制，抑制无用信息，提高网络的检测准确率。针对遥感图像观测面积大而目标相对较小，目标尺寸差异较大且分布不均匀的特点来说，降低了目标的误检率和漏检率。实验结果表明，改进的CenterNet算法相较于原始CenterNet算法的效果有明显提升，证明了改进算法的鲁棒性。

关键词: 目标检测, 遥感图像, 深度可分离卷积, 注意力机制, Anchor-Free机制, CenterNet算法