Pixel-Scale-Estimation-Based Instance Segmentation of Sensitive Targets in Tiled Remote Sensing Imagery

Security screening of tiled remote sensing imagery is challenging because tile boundaries fragment targets into multiple independent inputs, while missing spatial-resolution metadata makes it difficult to determine observation fields appropriate for different target pixel scales. Simply mosaicking adjacent tiles restores spatial continuity but does not eliminate secondary truncation caused by fixed-window partitioning. To address this issue, we propose a pixel-scale-guided instance-segmentation framework in which preliminary segmentation results are matched with a class-specific prior pixel-scale library to estimate image scale, and the estimated scale is used to adapt the sliding-window size and overlap ratio. To improve the representation of multiscale and fine-grained targets, Triplet Attention is integrated into YOLO11-Seg, while topology-constrained Hybrid Merge Non-Maximum Suppression (HM-NMS) associates cross-window masks according to overlap, boundary relationships, and spatial topology to recover unique complete instances. Experiments on The Mobile Military Target Dataset (MMTD) and two real-world tiled remote sensing datasets show that the proposed method achieves a precision of 1.000 on both test datasets, with recall values of 0.9795 and 0.9770, respectively. The results demonstrate that the proposed framework improves instance completeness and uniqueness while maintaining efficient inference under incomplete spatial-resolution metadata.