This study targets detection degradation in the pump-downstream channel （conveyor/dewatering chute） of light-luring purse-seine operations in the South China Sea， where fish exhibit small size， high speed， strong motion blur， and dense occlusion. A blur-tolerant， detail-preserving real-time detector， MGI-RTDETR， is presented. Built on RT-DETR， it integrates Multi-Scale Grouped Interaction features （MGI-FE）， Adaptive Dynamic Context Mixing （ADC-MI）， Adaptive Detail-Preserving Fusion （ADPF）， deployment-time Consolidated Re-parameterization （CRC）， and an Adaptive Reverse-Projection Upsampler （ARPU）， improving small-object boundary restoration and separability in crowded scenes without increasing inference-time complexity. A 396-image single-class dataset （train/val/test=7/1/2） is used under a unified 640×640 training/inference protocol. On the test set， the method attains Precision/Recall/F-Score of 88.54%/75.03%/81.23%， mAP50 of 81.00%， and mAP_50-95 of 28.29%， while maintaining deployment-friendly cost （44.5 GFLOPs， 13.39 M parameters，~50.46 FPS）， outperforming RT-DETR-R18 and multiple YOLO/SSD baselines. To illustrate downstream usability， ByteTrack is stacked without modifying the detector to realize line-crossing counting； counting precision （CP） at low/medium/high densities is 95.0%/80.8%/49.7%. To assess cross-vessel generalization， an external test set of 309 images from three additional vessels is evaluated， confirming trends consistent with the main experiment under varied viewpoints and illumination. The results indicate that multi-scale interaction and cross-scale adaptive fusion oriented to blur and fine-detail preservation provide a practical foundation for near-real-time， georeferenced catch analytics on working vessels， supporting transparency and data-driven fisheries management.