CSnNet: A Remote Sensing Detection Network Breaking the Second-Order Limitation of Transformers With Recursive Convolutions

In recent years, Transformer-based networks, known for their ability to model long-range dependencies, have been widely used in downstream computer vision tasks, surpassing certain neural network architectures. However, Transformer-based networks suffer from issues, such as large parameter size, high-computational complexity (CC), and difficulties in extending spatial and channel features to the third or even higher orders, resulting in convergence challenges for small- to medium-sized datasets and limited effectiveness in extracting high-order detailed features. In this article, we propose a high-order spatial and channel controllable convolution module, named
$\text {CS}^{n}$
, which can replace standard convolutions in any convolutional network. In the context of remote sensing small object detection,
$\text {CS}^{n}$
demonstrates superior performance compared with self-attention structures embedded in neural networks. Moreover, it introduces long-range dependency relationships among pixels, similar to self-attention, and adopts a cascaded recursive approach to extend spatial and channel features to arbitrary higher orders without introducing significant additional computation. This extension captures crucial information from high-order spatial and channel dimensions, resulting in improved accuracy for small object detection. In addition, we construct a novel, versatile
$\text {CS}^{n}$
-[feature pyramid network (FPN) + path aggregation network (PAN)] structure for object detection networks, referred to as
$\text {CS}^{n}$
Net. Finally, our proposed model exhibits significant advantages in remote sensing detection when compared with state-of-the-art methods on publicly available synthetic aperture radar (SAR) datasets [SAR ship detection dataset (SSDD) and high resolution SAR images dataset (HRSID)] and optical remote sensing dataset [Northwestern Polytechnical University very-high-resolution-10 dataset (NWPU-10)], achieving respective improvements of 3.4%, 4.5%, and 0.4% in
${\text {mAP}}_{{50}}$
compared with baseline models.