Efficient Multiaspect Compensation Network for Small Object Detection in SAR Images

Deep learning has demonstrated its potential capability in object detection of synthetic aperture radar (SAR) images. However, the detection of small objects still faces unique challenges due to insufficient detailed information and complex background clutter distribution. Although existing methods alleviate these issues to some extent by extracting discriminative features, the detection performance is still highly unsatisfactory due to the inevitable information loss as neural networks deepen. In this work, a multiaspect compensation network (MACNet) is proposed to improve the detection accuracy of small objects, in which two key components interactive encoding enhancement (IEE) module and multi-scale feature enhancement guided (MFEG) network are introduced. Specifically, IEE initially encodes the low-level details and highlevel semantics from neighboring features to enhance the feature representation of small objects. Besides, MFEG performs global modeling using Mamba to extract high-frequency information in SAR images, then integrates it with higher-level predicted features to compensate for information loss in multiscale features. Moreover, since the commonly used loss functions often exhibit insensitivity to small objects, a more effective scale and position sensitive (SPS) loss function is designed to improve attention to small objects. Numerical results on the SAR datasets LS-SSDD-V1.0, HRSID, and SAR-AIRcraft-1.0 demonstrate the effectiveness and superiority of the proposed strategy. The code is available at https://github.com/CongLi-18/MACNet.