SDCoNet: Saliency-Driven Multi-Task Collaborative Network for Remote Sensing Object Detection

TL;DR

SDCoNet tackles the challenge of small-object detection in low-quality remote sensing images by tightly coupling single-image super-resolution with detection through a cross-task shared Swin Transformer encoder. It introduces a saliency-driven query token mechanism and a gradient routing strategy to align SR reconstruction with detection semantics while focusing attention on object regions. The framework achieves state-of-the-art performance on NWPU VHR-10-Split, DOTAv1.5-Split, and HRSSD-Split, with pronounced gains on small objects and favorable computational efficiency. This approach demonstrates that explicit cross-task interaction and saliency-guided attention can significantly improve detection robustness in cluttered RS scenes, offering practical impact for Earth observation tasks.

Abstract

In remote sensing images, complex backgrounds, weak object signals, and small object scales make accurate detection particularly challenging, especially under low-quality imaging conditions. A common strategy is to integrate single-image super-resolution (SR) before detection; however, such serial pipelines often suffer from misaligned optimization objectives, feature redundancy, and a lack of effective interaction between SR and detection. To address these issues, we propose a Saliency-Driven multi-task Collaborative Network (SDCoNet) that couples SR and detection through implicit feature sharing while preserving task specificity. SDCoNet employs the swin transformer-based shared encoder, where hierarchical window-shifted self-attention supports cross-task feature collaboration and adaptively balances the trade-off between texture refinement and semantic representation. In addition, a multi-scale saliency prediction module produces importance scores to select key tokens, enabling focused attention on weak object regions, suppression of background clutter, and suppression of adverse features introduced by multi-task coupling. Furthermore, a gradient routing strategy is introduced to mitigate optimization conflicts. It first stabilizes detection semantics and subsequently routes SR gradients along a detection-oriented direction, enabling the framework to guide the SR branch to generate high-frequency details that are explicitly beneficial for detection. Experiments on public datasets, including NWPU VHR-10-Split, DOTAv1.5-Split, and HRSSD-Split, demonstrate that the proposed method, while maintaining competitive computational efficiency, significantly outperforms existing mainstream algorithms in small object detection on low-quality remote sensing images. Our code is available at https://github.com/qiruo-ya/SDCoNet.

Figures (7)

• Figure 1: Visualization of activation maps for different tasks. Fig. 1 (a) shows activation maps of the super-resolution task, which mainly attends to pixel-level texture and edge details, while Fig. 1 (c) shows activation maps of the object detection task, which primarily focuses on high-level semantic structures and object regions.
• Figure 2: The framework pipeline of SDCoNet. Our SDCoNet comprises two core branches: super-resolution and object detection, consisting of a shared encoder, a super-resolution decoder, a saliency-driven query filtering module, and an object detection encoder-decoder; in the first phase, the SR branch is frozen to update the object detection network, and in the second phase, dual-branch collaborative training is initiated.
• Figure 3: Example images from the DOTAv1.5-Split DOTA2018, NWPU VHR-10-Split Cheng2016, and HRSSD-Split Zhang2019 datasets. The left side shows high-resolution images from these datasets, while the right side presents the corresponding low-resolution images generated by bicubic downsampling.
• Figure 4: Category distribution of the benchmark datasets. The pie charts show the class proportions of the DOTAv1.5-Split DOTA2018, NWPU VHR-10-Split Cheng2016, and HRSSD-Split Zhang2019 datasets, which contain a large number of small-object categories such as storage-tank, small-vehicle, bridge, and ship.
• Figure 5: Per-class detection performance on the NWPU VHR-10-Split dataset. The chart compares SDCoNet with representative CNN-based (Faster R-CNN, YOLOX-s, FFCA-YOLO, LEGNet-T) and Transformer-based (DINO, DN-DETR) detectors, showing consistent gains for small-object categories such as storage-tank and vehicle.