| 摘 要: 遥感图像目标检测因目标类型多样、尺寸微小、分布密集、背景杂乱等特点,仍面临现有深度学习模型难以有效利用背景信息、微小目标检测效果欠佳的挑战。针对此问题,本文提出基于抑扰跨注意与逆残差融合的遥感微目标检测网络(PCINet)：设计C3S抑扰跨注意模块实现多轴向特征深度交叉融合,以高效捕获全局上下文信息；在主干网络引入部分卷积前馈网络,增强微小目标表征能力；构建逆残差多尺度自适应融合块,提升网络复杂场景适应能力。在DOTA、HRSC2016和UCAS-AOD数据集上的实验表明,该方法mAP分别达77.84%、92.23% 和90.42%,相较此前最优方法提升0.64%、0.53%和0.32%,性能显著优于主流检测算法。 |
| 关键词: 深度学习 人工智能 遥感图像 小目标检测 |
|
中图分类号: TP
文献标识码:
|
|
| PCINet: A Remote Sensing Tiny Object Detection Network Integrating Suppressive Cross-Attention and Inverse Residual Fusion |
|
Wang Honglin1, Jia Yuwei2
|
1.School of Artificial Intelligence, Nanjing University of Information Science and Technology;2.School of Computer Science and School of Cyberspace Security, Nanjing University of Information Science and Technology
|
| Abstract: Remote sensing image object detection still faces challenges due to the diverse types of objects,their small sizes,dense distribution,and cluttered backgrounds,which make it difficult for existing deep learning models to effectively utilize background information and achieve satisfactory detection results for small objects.To address this issue,this paper proposes a remote sensing small object detection network(PCINet)based on suppressed cross-attention and inverse residual fusion:a C3S suppressed cross-attention module is designed to achieve deep cross-fusion of multi-axis features, efficiently capturing global contextual information;a partial convolution feedforward network is introduced into the backbone network to enhance the representation ability of small objects;and an inverse residual multi-scale adaptive fusion block is constructed to improve the network"s adaptability to complex scenes.Experiments on the DOTA,HRSC2016,and UCAS-AOD datasets show that the method achieves mAP of 77.84%,92.23%,and 90.42%,respectively,which is 0.64%,0.53%,and 0.32% higher than the previous best method,demonstrating significantly superior performance compared to mainstream detection algorithms |
| Keywords: deep learning Artificial Intelligence Remote sensing image small object detection |
