Abstract
Synthetic Aperture Radar (SAR) has been commonly utilized in both military and civilian fields in the past decade because of its all-weather, high-resolution, and strong penetration capabilities. However, due to its unique coherent imaging characteristics, speckle noise will be generated in SAR images, which greatly degrades the image quality and hinders subsequent target recognition work. In this paper, several denoising methods will be compared and estimated by indexes such as signal-to-noise ratio (SNR) and other indicators to estimate the model quality. Then the better method will be selected to preprocess SAR images and perform target recognition on the processed images. Based on the classic CNN model, a Squeeze-and-Excitation block will be added into the convolutional layer and train the model using part of the MSTAR dataset SAR images. During training, the SAR image samples are preliminarily divided into 8 classes, and the neural network is trained using the training set and validation set. When the training accuracy, validation accuracy, and validation loss value reach a relatively ideal value, training is stopped. Finally, the model is used to recognize the SAR image categories and compare the average recognition accuracy with the actual image categories. In addition, by comparing the average recognition accuracy before and after noise suppression, the experiments can also verify the performance of the method in SAR image noise suppression.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call