Ship Classification Based on Sidelobe Elimination of SAR Images Supervised by Visual Model

SAR image of a ship at the ocean usually contains severe cross sidelobe due to the strong scatter points of the ship body, which increases the difficulty of detecting the ship type correctly. Aiming at this problem, this paper proposes a novel method to eliminate the sidelobe of the strong scattering points in the ship's original SAR image, which utilizes the contour of the ship visual model to extract the ship hull information. We first estimate the ship's aspect angle in the original SAR image and transmit this parameter into the top view dictionary of the ship visual model to find out the top view image with the same aspect angle. Then, we extract the ship visual model's contour to merge with the original SAR image, preserving the ship body's grayscale information in the visual model contour and rejecting the rest grayscale information outline. Finally, we construct a convolutional neural network with a simple structure to train the newly generated SAR image to finish the ship classification task. To validate our proposed method's feasibility, we conduct the ship classification experiment on the latest benchmark SAR ship dataset. The proposed method achieves the final average classification accuracy of 98.21%, far exceeding the original method provided in this dataset.