Investigation of the Influence of Speckle Noise on the Accuracy of Object Detection by Convolutional Neural Networks

Abstract

Synthetic aperture radars are widely used to obtain high-resolution radar images without the influence of weather and lighting conditions, at large distances and in a wide swath. Object detection and classification is one of the main tasks in radar image processing. One of the ways to do this is to use convolutional neural networks. However, neural network training requires a large set of images (dataset). There are few radar image datasets in the public domain and they are not labeled. This article discusses the possibility of using convolutional neural networks trained on optical images for detecting and classifying objects in radar images. Optical and radar images have both similarities and differences. One important feature of SAR images is a high level of speckle noise that is not present in optical images and must be reduced. Therefore, the influence of speckle noise on the detection accuracy is investigated.