MLC-Net: A Robust SAR Ship Detector With Speckle Noise and Multiscale Targets

Ship detection is a critical component of marine resource management and environmental monitoring. Methods based on deep learning have been widely used in target detection. However, there are two obstacles in ship detection: First, due to the mechanism of synthetic aperture radar (SAR) imaging, SAR images contain a lot of speckle noise, which affects the accuracy of detection; second, ships in the image have multiscale characteristics caused by significant differences in ships' sizes and imaging resolutions. As convolutional neural networks (CNNs) become deeper, the features of small targets gradually diminish or disappear. Therefore, the network is more inclined to extract the features of large targets, resulting in missed and false detection. To solve the above problems, we propose a one-stage ship detection network for speckle noise and multiscale targets called MLC-Net. First, to mitigate speckle noise, we introduce a simplified morphological denoising method that reduces the computational burden of the algorithm. Second, we present a large separable kernel module to expand the receptive field and create a new backbone network, ML-CSPDarknet, addressing multiscale target detection. Finally, we propose cascade adaptively channel-spatial feature fusion, which effectively fuses features from the backbone at all levels to enhance network performance further. Experimental results on the SAR ship detection dataset and the high-resolution SAR images dataset demonstrate that the detector achieves state-of-the-art performance, with AP reaching 99.16% and 93.08%, respectively. Furthermore, experiments on a large-scale SAR image show that the method has great migration application ability.