67 / 2024-08-14 18:02:45

SKSH-Net: Scattering Keypoint Structural Hints-Based Network for Aircraft Detection in SAR Images

SAR aircraft detection, scattering keypoint, scattering topology

Aircraft detection in synthetic aperture radar (SAR) imagery is significant due to its critical role in various applications, including surveillance, reconnaissance, and security. To meet the demands of production and daily life, SAR aircraft detection algorithms have been evolving rapidly over the past decades, transitioning through two major phases: traditional methods and deep learning-based methods. Traditional methods are primarily classified into two categories: grayscale-based models and geometrical-based models. By modeling the background, constant false alarm rate (CFAR), a typical representative of grayscale-based models, obtains the threshold used to distinguish targets from the background under the condition of constant false alarm rate. As the resolution becomes finer, more geometric features of the target are presented in the SAR images, which supports the advancement of geometrical-based model. With abundant SAR images are more readily available, creating an auspicious environment to develop deep learning on SAR images. As one of the initial attempts, auto-encoder extracts features from raw images to train a soft-max classifier. With this simple combination, deep learning alleviates the reliance of hand-crafted features. And some general and powerful modules are also present in the models. For example, the feature pyramid module improves the performance of detecting aircraft at different scales. These methods represent an extension of deep learning paradigms and have demonstrated superior performance.

However, in view of the characteristics of SAR images, certain challenges remain and warrant focused attention for effective SAR aircraft detection. The wavelength of the SAR signal is longer than the wavelength of visible light, and the observed targets exhibit discrete and multi-centric characteristics. In particular, artificial facilities such as corridors and bridges in complex airport environments demonstrate similar characteristics. In addition, inherent factors such as speckle noise and sidelobes inevitably emerge, which tend to blur target edge contours and weaken target texture information. Overall, aircraft features in SAR images are discrete and susceptible to interference from factors such as complex backgrounds and noise. However, most of the existing detection methods uniformly focus on the aircraft region and exhibit limited capability in associating discrete and fragmented target features. This limitation exacerbates the difficulty in accurately mapping multiple scattering centers to a single aircraft, consequently constraining the upper limit of performance.

To solve the above problems, an innovative scattering keypoint structural hints-based network (SKSH-Net) is proposed for aircraft detection in SAR images. The overall structure of the SKSH-Net is shown in Fig. 1, comprising three main components: the backbone with FPN, the star topology feature fusion module (STFFM) and the completeness-aware detection head. SKSH-Net constructs a star topology (ST) by connecting discrete scattering centers according to spatial relationships. Specifically, the star topology space fusion module (STSFM) utilizes the neighboring information of the ST to dynamically aggregate the discrete scattering centers. The star topology channel attention module (STCAM) leverages node information of the ST to enhance aircraft saliency in the presence of strong interference. Additionally, a completeness and consistency loss (CCLoss) is formulated for aircraft with discrete, multiple scattering centers. Without limiting itself to continuous and complete targets, CCLoss can guide the model to obtain predictions that more perfectly cover the ST, which means more accurate aircraft localization.