Object detection in low-resolution optical-SAR remote sensing imagery is challenged by degraded object structures, residual cross-modal spatial inconsistency, and unreliable SAR responses. To address these issues, we propose TPA-NET, a progressive cross-modal alignment and confidence fusion network built on YOLOv10. TPA-NET adopts a three-stage coarse-to-fine processing strategy. The Shallow Coarse Alignment (SCA) module performs initial feature-space calibration, the Pyramid Residual Alignment (PRA) module further refines local cross-modal correspondence across multiple scales, and the Detection-aware Confidence Fusion (DCF) module dynamically regulates modality contributions according to object responses and task-driven feature reliability. The framework is trained end-to-end without additional image-level registration or manual correction. Experimental results show that TPA-NET achieves 95.4% mAP50 and 71.4% mAP50–95 on OGSOD 1.0, improving mAP50–95 by 1.9 percentage points over IRDFusion. On M4-SAR, it achieves 86.1% mAP50 and 62.0% mAP50–95, exceeding E2E-OSDet by 0.5 and 0.8 percentage points, respectively. Ablation and visualization analyses further support the effectiveness of progressive feature calibration and reliability-aware fusion for improving cross-modal object localization.