SAR Target Classification Based on Radar Image Luminance Analysis by Deep Learning
Hongliang Zhu, Weiye Wang, Rocky Leung
Abstract
This letter introduces a novel synthesis aperture radar (SAR) target classification method that is quite different from the conventional ones by using the luminance information of the radar image. With the development of deep learning, increasingly more image classification problems are adopted with this kind of popular technology, as is the SAR classification issue. After reading, most articles refer to this problem using the deep learning method; we find most of them just regard the raw data as an input to the neural networks and output the final classification result with several steps. The raw SAR data are also very limited, which often requires data augmentation in the deep learning method. Besides this, speckle noise in a SAR image is also a severe problem to discriminate the target preciously. To solve this problem, our approach extracts the luminance information of each SAR image and forms the target's profile with the luminance level, which reduces the speckle noise significantly. After that, we construct a convolutional neural network to train these luminance outline images instead of the raw SAR images. The experimental results on the MSTAR dataset indicate our approach that achieves a higher accuracy rate than the other state-of-the-art methods, which is close to 100% in each category.