AMDC-YOLO: An Adaptive Multi-Dimensional Dynamic Convolution Approach for Aircraft Target Detection

With the rapid development of deep learning, remarkable breakthroughs have been achieved in SAR image target detection based on convolutional neural networks. However, the inherent noise in SAR images and the low contrast between targets and backgrounds still pose challenges to detection accuracy [1]. To address these issues, this paper proposes an improved YOLO algorithm for aircraft target detection in SAR images. This algorithm integrates the BiFormer module and the DynamicConv module for adaptive multi-feature fusion. Firstly, the BiFormer module is introduced. Its unique bidirectional feature fusion ability can automatically adjust feature sampling, effectively capturing target features of different shapes and positions. Secondly, the DynamicConv module replaces traditional convolution and pooling operations, reducing feature information loss and enhancing the feature extraction ability for small targets and low-resolution images. Additionally, a novel bilayer routing attention mechanism is applied. It strengthens the effective fusion of blurred and local features through a dynamic sparse attention mechanism. The global feature flow of the BiFormer structure further improves the detection accuracy of small targets. Experiments on the SAR-Aircraft-1.0 dataset show that the proposed method significantly outperforms existing advanced methods in terms of target detection performance.