基于YOLOX的遥感飞机检测算法

Remote Sensing Aircraft Detection Algorithm Based on YOLOX

摘要: 针对遥感飞机检测中存在背景信息复杂、目标尺寸小、特征不明显、多尺度等问题，提出了一种基于YOLOX的遥感飞机目标检测算法。本文以YOLOX模型为基准，在残差网络的基础上加入通道注意力和空间注意力，使网络专注于学习图像中具体目标的结构而忽略背景图的内容细节，从而提高了目标检测的精度，并且对小目标的检测能力、特征不明显的目标的检测精度也有一定的提升；同时，用空洞空间金字塔池化(At-rous Space Pyramid Pooling, ASPP)替换空间金字塔池化(Space Pyramid Pooling, SPP)，ASPP可以聚合多尺度上下文信息，增大感受野，提高模型检测同一类别不同尺寸的能力。实验结果显示，在RSOD数据集上AP@0.5提高了1.17%，在UCAS-AOD数据集上AP@0.5提高了1.03%。

Abstract: A YOLOX-based algorithm for remote sensing aircraft target detection is proposed for the problems of complex background information, small target size, inconspicuous features, and multi-scale in remote sensing aircraft detection. In this paper, the YOLOX model is used as a benchmark, and channel attention and spatial attention are added to the residual network, so that the network focuses on learning the structure of specific targets in the image and ignores the content details of the background map, thus improving the accuracy of target detection, and the detection capability of small targets and the detection accuracy of targets with obscure features; at the same time, atrous space pyramid pooling (At The difference between ASPP and SPP lies in the fact that ASPP incorporates cavity convolution with a larger field of perception for the features extracted from the image and allows the resolution to be aggregated without too much degradation, allowing the aggregation of multi-scale contextual information and increasing the field of perception to improve the model. information, increase the perceptual field and improve the model’s ability to detect different sizes of the same category. The experimental results show that on the RSOD dataset AP@0.5 increased by 1.17%, and on the UCAS-AOD dataset AP@0.5 increased by 1.03%.