Remote Sensing Target Detection Based on Information Expansion Collection and Adaptive Feature Fusion

在现有方法的基础上，本文提出了一种新颖的信息扩大集合和自适应特征融合的检测算法。文中在主干网络部分引入ConvNeXt模块来加强对被遮蔽目标的检测能力，提出了信息扩大集合模块来充分地利用图像中的上下文信息，优化对长宽比较大目标的检测效果。使用协调注意力模块来防止目标位置信息的丢失，通过挤压与激励模块对通道重新进行权重分配，挑选出重要性较高的通道进行计算。文中还使用自适应空间特征融合模块，对不同层级的特征图进行融合来保证金字塔的效果。在DOTA-v1.5遥感数据集上，相较于原始网络，文中方法mAP@0.5性能提升了2.5个百分点。在另外的2个数据集上，本文提出的算法也取得了更好的检测性能。On the basis of existing methods, this paper proposes a novel detection algorithm based on Information Expansion Collection and Adaptive Feature Fusion. The ConvNeXt module is introduced in the backbone network part to enhance the detection ability of occluded targets; the Information Expansion Collection module is proposed to fully utilize the contextual information in the image to optimize the detection effect of targets with large aspect ratio; and the Coordinated Attention module is used to prevent the loss of target position information. The Squeezing and Excitation module is used to re-assign weights to the channels and select the channels with higher importance for computation; the Adaptive Spatial Feature Fusion module is used to fuse the feature maps of different layers to ensure the effect of pyramid. Finally, on the DOTA-v1.5 remote sensing dataset, compared with the original network, the mAP@0.5 performance improvement of 2.5% was obtained. The algorithm proposed in this paper also achieved better detection performance compared to the current state-of-the-art detection algorithms on the other 2 datasets.