Despite significant challenges, the text-guided remote sensing image generation method shows great potential in many practical applications such as generative adversarial networks in remote sensing tasks; generated images still face challenges such as low realism, face challenges, and unclear details. Moreover, the inherent spatial complexity of remote sensing images and the limited scale of publicly available datasets make it particularly challenging to generate high-quality remote sensing images from text descriptions. To address these challenges, this paper proposes the RSVQ-Diffusion model for remote sensing image generation, achieving high-quality text-to-remote-sensing image generation applicable for target detection, simulation, and other fields. Specifically, this paper designs a spatial position encoding mechanism to integrate the spatial information of remote sensing images during model training. Additionally, the Transformer module is improved by incorporating a short-sequence local perception mechanism into the diffusion image decoder, addressing issues of unclear details and regional distortions in generated remote sensing images. Compared with the VQ-Diffusion model, our proposed model achieves significant improvements in the Fréchet Inception Distance (FID), the Inception Score (IS), and the text–image alignment (Contrastive Language-Image Pre-training, CLIP) scores. The FID score successfully decreased from 96.68 to 90.36; the CLIP score increased from 26.92 to 27.22, and the IS increased from 7.11 to 7.24.

Multi-modal remote sensing object detection faces challenges of significant nonlinear imaging discrepancies and semantic feature misalignment between optical and Synthetic Aperture Radar (SAR) imagery. Existing fusion algorithms often struggle to effectively distinguish complementary information from modality-specific noise, leading to feature redundancy or even negative transfer. To address these issues, this paper proposes a feature-level fusion strategy based on correlation-guided progressive interaction and gated fusion, which is embedded into a single-stage detection network to enhance multi-modal detection performance. For the feature enhancement stage, a Correlation-Guided Progressive Spatial Interaction (CGPSI) module is designed. This module utilizes Cross-Modal Spatial Correlation Modeling (CMSCM) to construct guidance priors and couples them with a Progressive Spatial Modulation (PSM) mechanism to generate spatial weight maps that dynamically evolve with the training process. Simultaneously, bidirectional parallel Attention-based Cross-Modal Retrieval (ACMR) is employed to capture complementary contextual features across a global receptive field. Subsequently, the generated dynamic weights are utilized to perform fine-grained spatial weighting and filtering, accurately extracting discriminative gains and suppressing cross-modal noise interference. Finally, the modulated complementary information is integrated into the backbone path through residual connections to achieve the deep integration of complementary information, significantly strengthening the representation capability of target features. For the feature fusion stage, a Gated Weight Fusion (GWF) module is adopted to identify discriminative differences in dual-modal features, effectively suppressing background noise and clutter interference in complex scenes by generating adaptive spatial weight maps. Experimental results on the OGSOD (Optical Image Guided SAR Image Object Detection) dataset demonstrate that the proposed method achieves an $mAP_{50}$ of 0.873, significantly outperforming existing multi-modal detection algorithms. Ablation studies further confirm the synergistic superiority of the proposed modules: the CGPSI module enhances target feature intensity through deep interaction, while the GWF module achieves superior false alarm suppression through precise gain allocation, collectively ensuring the detection robustness of the model in heterogeneous remote sensing scenarios. Our code is accessible at https://github.com/ciomphzx/DualBranch-OptSAR-YOLO.

To address the attenuation of shallow fine-grained structural information during deep feature propagation and the representational conflict between classification and regression in remote sensing tiny object detection, this paper proposes a dual asymmetric detection framework, termed $A^2$-Det. Built upon YOLO11n, the detection pyramid is shifted toward the high-resolution P2–P4 levels by introducing a P2 detection branch and removing the original P5 stage, thereby reducing the loss of fine-grained spatial details caused by successive downsampling. On this high-resolution feature basis, a Query–Key–Value (QKV)-guided Asymmetric Spatial Feature Enhancement module (Q-ASFE) is deployed at the P2 and P3 stages, where direction-sensitive asymmetric convolutions are combined with lightweight QKV-based contextual modulation to strengthen tiny object structural responses while suppressing complex background interference. Furthermore, an Asymmetric Coordinate–Semantic Decoupled Head (ACS-Head) performs differentiated modeling for the semantic selection required by classification and the coordinate-sensitive representation required by regression, thereby alleviating task-specific representational conflict. Through the progressive coordination of high-resolution feature preservation, shallow-feature enhancement, and task-specific prediction, $A^2$-Det improves mAP₅₀ by 5.8 percentage points over YOLO11n on the VisDrone dataset. Consistent improvements on the USOD and RSOD datasets further demonstrate its effectiveness and cross-scene generalization.