HRS-Former: A High-Resolution Ship Remote Sensing Image Feature Extraction Model

In recent times, the application of remote sensing imagery has been expanding, particularly in the identification of maritime ships. In complex backgrounds, the utilization of Transformers at the feature level is becoming increasingly widespread. However, traditional Transformers may lead to loss and simulate multiscale variations during the feature extraction process. To address these issues, a High-Resolution Special Transformer (HRS-Former) is proposed based on the Swin Transformer, incorporating rich spatial coordinates. Additionally, a dense prediction approach for high-resolution representation is constructed, thereby enhancing the image's feature extraction capability. Firstly, to fully leverage the advantages of global information in the Local Transformer, a Global Attention is introduced between various stages of the Swin Transformer. Secondly, a multi-scale parallel cross-design from the High-Resolution Convolutional Network is incorporated. The experimental results indicate a respective increase of 12.7% and 3.2% in mean average precision (mAP) on the remote sensing ship datasets HRSC2016 and FGSD2021, compared to the baseline methods.