RSTR: An Efficient Detection Network with Cross-Attention Multiscale Feature Aggregation for Remote Sensing Applications

Remote sensing scenes exhibit severe scale imbalance, dense layouts, and heavy background clutter, often reducing objects of interest to only a few pixels. Existing detectors typically exhibit a high miss rate under these challenging conditions. We present RSTR (Remote Sensing Detection Transformer), a transformer-based framework optimized for tiny-object detection in complex backgrounds. Specifically, we introduce a novel cross-attention multi-scale feature pyramid network (CMFPN) that captures cross-scale interactions and fuses high-level semantics with low-level details to strengthen representations of minute targets. A deformable patch-wise attention (DPA) module links global context and local detail to increase attention to tiny objects across channels and spatial positions, improving detectability. A lightweight FocusBlock accelerates inference while preserving cues for tiny targets and maintaining a small parameter budget, achieving a favorable speed-accuracy trade-off. We further design FMIoU, a loss function that improves learning from low-confidence predictions and reduces false positives and missed detections for tiny objects. Relative to comparable baselines, RSTR reduces the number of parameters by 39.8% and computation by 16.2%, indicating a stronger accuracy-efficiency balance. Extensive experiments demonstrate that RSTR achieves AP50 scores of 51.3%, 58.8%, and 78.0% on the VisDrone, AI-TOD, and DIOR datasets, respectively. RSTR also yields substantial gains over the RT-DETR baseline and other detectors on tiny-object benchmarks, providing an effective solution for tiny object detection in remote sensing imagery.