Multi-scale fusion method for target detection in remote sensing images
Lujianheng1  Liujunnuo1  Zhanglu1  Wangning2（Corresponding author）

1 School of Artificial Intelligence, Guangzhou Huashang University, Guangzhou, Guangdong, 511300；

2 School of Artificial Intelligence, Guangzhou Huashang University, Guangzhou, Guangdong, 511300；

Abstract : To address the challenges of scale changes and complex backgrounds in remote sensing images, this study designed an object detection approach using weighted multi-scale feature fusion. This approach was optimized on the CSPDarknet backbone, using dilated convolutions to widen the receptive field, deformable convolutions for spatial alignment, and a channel attention mechanism to dynamically adjust feature weights. Ultimately, the detection function is implemented using a three-branch decoupled detection head. Experimental results in DOTA v2.0 indicate that this approach outperforms existing mainstream models in terms of accuracy, recall, and response speed, and demonstrates excellent robustness and universal adaptability in cross-domain testing.

Keywords: multi-scale fusion; remote sensing imagery; object detection; small targets; real-time performance

References

[1]Xinnan FAN, Wei YAN, Pengfei SHI, et al. Remote sensing image target detection based on a multi-scale deep feature fusion network[J]. National Remote Sensing Bulletin, 2022, 26(11): 2292-2303.

[2]Wang Y, Tian Y, Liu J, et al. Multi-stage multi-scale local feature fusion for infrared small target detection[J]. Remote Sensing, 2023, 15(18): 4506.

[3]Huang W, Li G, Chen Q, et al. CF2PN: A cross-scale feature fusion pyramid network based remote sensing target detection[J]. Remote Sensing, 2021, 13(5): 847.

[4]Xu H, Zhong S, Zhang T, et al. Multiscale multilevel residual feature fusion for real-time infrared small target detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1-16.

[5]Du Z, Liang Y. Object detection of remote sensing image based on multi-scale feature fusion and attention mechanism[J]. IEEE Access, 2024, 12: 8619-8632.