MKD-Net: A Multikernel Downsampling Network for Adjacent and Small-Scale Ship Detection in SAR Images

Abstract

Synthetic aperture radar (SAR), as an important remote sensing technology, is playing an increasingly important role in marine monitoring. Nevertheless, in ship detection tasks of SAR images involving complex near-shore backgrounds, blurred edges of adjacent targets, and dense small ships, false alarms and missed detections remain difficult to avoid. To this end, this letter proposes a ship detection network, termed MKD-Net, which is based on a multikernel downsampling (MKD) module. This method integrates two key designs with low-parameter complexity. The MKD transfers spatial information into the channel domain and extracts information using multikernel convolution, which can obtain features under different receptive fields while preserving fine-grained image details. The content-adaptive coordinate attention (CACA) introduces a dual-layer gating mechanism that consists of basic learnable weights and content-adaptive weights. Compared with other methods, the proposed method uses only 2.9M parameters. Furthermore, it can effectively detect adjacent targets in complex backgrounds, as well as dense small-scale targets. The proposed method achieves optimal detection performance.