Local Constraints Convolutional Neural Network for SAR Image Denoising and Target Configuration Recognition

Synthetic aperture radar (SAR) target recognition is an important branch of SAR image processing. To overcome the influences of inevitable speckle noise, especially for similar configurations recognition, we propose a local constraints convolutional neural network (LC-CNN) for joint SAR image denoising and target configurations recognition. The proposed LC-CNN enhances recognition performance through a collaboratively designed multitask loss function. In the denoising stage, a speckle suppression loss is designed to smooth background noise whereas retaining target details. In the recognition stage, a local structure maintenance loss is designed to enhance discrimination of similar configurations by maintaining local geometric relationships. And a feature invariance loss is established to ensure core target features remain stable after denoising. Experimental results demonstrate LC-CNN's robustness under varying speckle noise levels and excellent performance in similar SAR target configurations recognition.