Remote sensing imagery (RSI) object detection is critical to many applications, yet mainstream detectors analyse only spatial features and, because of spectral bias, fail to learn high-frequency information adequately, resulting in performance bottlenecks under cluttered backgrounds, distractors, and multi-scale targets, especially small ones. To break these limitations, we propose MSRS-DETR, an end-to-end framework that deeply fuses spatial and frequency cues. The approach introduces three key innovations: (1) C2f_FATNET, a frequency-attention-enhanced lightweight residual backbone that provides richer dual-domain features with fewer parameters; (2) an Entanglement Transformer Block (ETB) in the encoder that refines deep semantics via cross-domain frequency–spatial interaction and suppresses background interference; and (3) S2-CCFF, a shallow-feature-extended bidirectional fusion path that markedly improves the retention and utilisation of fine details for small objects. Experiments on HRSC2016 and ShipRSImageNet demonstrate the effectiveness and generalisation of this spatial–frequency paradigm: relative to the baseline, MSRS-DETR reduces parameters by $29.1\%$, boosts inference speed by $12.4\%$ and $8.4\%$, and raises mAP_50-95 by $1.69\%$ and $2.16\%$, respectively.

Small target detection in remote sensing images faces challenges due to complex backgrounds, weak features, and large scale differences. This paper proposes an improved YOLOv5-based network, termed ClearSight-RS, with the full name “Clear and Accurate Small-target Insight for Remote Sensing”. As the name implies, the network is dedicated to achieving clear feature perception and accurate target localization for small targets in remote sensing images. The improvements focus on three aspects: integrating an improved Dynamic Snake Convolution (DSConv) module into the backbone network to strengthen the extraction of small target boundaries and geometric features, as well as the expression of weak textures; embedding a Bi-Level Routing Attention (BRA) module in the Neck part to enhance target focusing and suppress background interference; and optimizing the detection head by retaining only shallow high-resolution feature layers for prediction, reducing feature loss and redundant computations. Experimental results show that, based on the VEDAI dataset, ClearSight-RS achieves the highest mAP for all 8 vehicle categories; based on the NWPU VHR-10 dataset, its overall mAP reaches 93.8%, significantly outperforming algorithms such as Faster RCNN and YOLOv5l; based on the DOTA dataset, the capability of the proposed BRA module in suppressing background interference and capturing small target features is demonstrated. The network balances accuracy and efficiency, performing prominently in detecting vehicles and multi-category small targets in complex backgrounds, verifying its effectiveness.

Remote sensing (RS) images are often degraded by atmospheric haze, which compromises both visual interpretation and downstream applications. To address this, we introduce CSGL-Former, a novel Cross-Stripes Global–Local Fusion Transformer for RS image dehazing. Our model efficiently captures anisotropic long-range dependencies using cross-stripes attention (CSA) and aggregates hierarchical global semantics via a Multi-Layer Global Aggregation (MLGA) module. In the decoder, global context is adaptively blended with fine-grained local features to restore intricate textures. Finally, inspired by the atmospheric scattering model, a soft reconstruction head restores the clear image by predicting spatially varying affine parameters, strictly preserving content fidelity while effectively removing haze. Trained end-to-end, CSGL-Former demonstrates a compelling balance of accuracy and efficiency. Extensive experiments on the RRSHID and SateHaze1K benchmarks show that our model achieves state-of-the-art or highly competitive performance against representative baselines. Ablation studies further validate the effectiveness of each proposed component.