Ship object detection in remote sensing images based on improved real-time detection transformer

Abstract

To address the challenges in detecting ship targets in offshore port remote sensing images—such as dense object distribution, complex backgrounds, and variable scales, which lead to difficulties for traditional methods and high computational costs—this study proposes an improved algorithm based on RT-DETR. This algorithm introduces three key modules: RepStem, LWGABlock, and AIFI-LRSA. The RepStem module replaces the two convolutional layers in the backbone network, reducing model complexity while maintaining feature extraction capability. The LWGABlock module enhances the detection performance for small targets by lowering the complexity of feature extraction and aggregating multi-scale outputs. The AIFI-LRSA module significantly reduces computational load and memory consumption by substituting the standard multi-head self-attention mechanism. Experimental resultsshow that the proposed model achieves a 4.7% improvement in mAP50 on the DOTA dataset, with its computational cost reduced to 40.2 GFLOPs, thereby significantly enhancing both detection efficiency and accuracy.