LSDNet: Lightweight CNN Model Driven by PNF for PolSAR Image Ship Detection

Abstract

Polarimetric synthetic aperture radar (PolSAR) as an active microwave imaging device employing pulse compression technology to obtain echo information of multiple polarization channels has been widely used in marine target detection. However, the detection of ship targets in PolSAR images is often disturbed by clutter interference, such as side lobes, ghost ships, etc. Additionally, the current common convolutional neural network object detection models are accompanied by a large number of parameters, thereby, their performance is greatly limited by available datasets. Therefore, aiming at solving the above questions, we propose a lightweight ship detection model driven by a polarimetric notch filter (PNF) feature for PolSAR ship detection. First, considering the scattering mechanism of target and clutter, the PNF feature is constructed with geometric perturbation information to overcome the vulnerability of existing amplitude feature-driven-based detectors to clutter interferences. Then, a lightweight ship detection network (LSDNet) is proposed, which uses a limited multilayer convolutional structure supported by a small-scale dataset and introduces dilated convolution during the extraction of high-level features to achieve spatial feature perception of weak and small ship targets. Finally, combined with the region-based fully convolutional head network, the model parameter is further reduced while training and inference efficiency is accelerated. Experimental results on a substantial number of PolSAR images show that the PNF-driven LSDNet model is better than the state-of-the-art detection models in terms of accuracy and efficiency, whose parameter quantity is 41.04% less than that of Faster R-CNN and achieves 0.92 of F1 Score and 0.9683 of average precision.