Scattering Speckle Information Extraction Network for Aircraft Detection in SAR Images

Abstract

Accurate detection of aircrafts in synthetic aperture radar (SAR) images has an important role in both military and civilian areas. However, aircrafts in SAR images are composed of several scattering speckles whose intensity and distribution are sensitive to the environmental factors and imaging conditions. Existing detection methods have difficulties capturing variable information of these scattering speckles. To alleviate this problem, a scattering speckle information extraction network (SSIEN) based on deep learning is proposed, which is capable of adaptively extracting speckle information of the targets for aircraft detection in various scenarios. In SSIEN, an information enhancement and extraction module (IEEM) is proposed. IEEM integrates two key components, weakly supervised deformable convolution module (WSDCM) and convolution block attention module (CBAM). The former is proposed to locate the speckles and extract the information, while the latter is employed to highlight valuable information and suppress interference. The experimental results on Gaofen-3 SAR images demonstrate the excellent performance of the proposed network for SAR image aircraft detection.