<p>Paper containers, as a critical packaging medium for fast-moving consumer goods such as coffee, are utilized in massive volumes. In the production process, achieving high-speed and high-precision identification of their surface defects is a crucial link to ensuring production efficiency, product quality, and food safety. However, identifying and eliminating tiny defects under the stringent real-time constraints of high-speed production lines is a technical bottleneck that severely restricts product yield rates. This paper is dedicated to lowering the real-time missed detection rate and proposes the Interaction-driven Multi-scale YOLO (IM-YOLO). IM-YOLO designs an Internal Feature Enhancement Network that balances local and global perception, and introduces the efficient Multi-Scale Shuffle Attention to enhance sensitivity to weak features. Furthermore, we have constructed and released the Multi-scene Paper Container Defect Dataset. Experimental results demonstrate that IM-YOLO achieves 83.2% mAP50 and 47.5% mAP50-95 while maintaining a real-time speed of 212.76 FPS, significantly outperforming other models. Most critically, the detection precision for the most easily missed spot and stain defects improved by 4.1 and 6.2 percentage points, respectively, providing powerful evidence of its effectiveness in lowering the missed detection rate. Generalization experiments further substantiated its superior performance. In conclusion, IM-YOLO provides a high-performance, low-missed-detection, real-time solution for industrial paper container quality inspection.</p>