A lightweight dual-domain attention YOLO for small object detection in remote sensing images

Small object detection (SOD) in remote sensing images (RSIs) remains challenging due to limited pixel representation, complex background interference, and the requirement for computationally efficient lightweight models. To address these issues, this paper proposes LDA-YOLO, a lightweight detection framework built upon YOLOv8 that integrates adaptive refinement of multi-level features and dual-domain attention to enhance fine-grained perception while maintaining low computational cost. Specifically, we design a Selective Gated Convolution (SGConv) module that adopts a multi-branch convolutional structure with ECA-based channel reweighting to selectively emphasize informative feature responses while suppressing redundant activation. An Improved C2f (IC2f) block is further constructed to strengthen short-range feature reuse and preserve detailed structures during feature refinement. To further refine the multi-level features aggregated by the neck, a Dynamic Convolutional-attention Fusion Module (DCFM) combines lightweight attention with depthwise convolutional enhancement to integrate long-range contextual cues and local structural features with complementary receptive fields. In addition, a Dual-Domain Attention Module (DDAM) is introduced in the prediction stage, where spatial- and frequency-domain responses are jointly modeled to highlight small-object structures and suppress cluttered backgrounds. Extensive experiments on three widely used remote sensing benchmarks, including VisDrone2019, DIOR and DOTA, demonstrate that LDA-YOLO achieves competitive SOD accuracy compared with existing lightweight detectors, while maintaining a compact model size and low computational cost, indicating its potential for resource-constrained deployment.