<p>Identifying out-of-distribution (OOD) samples is crucial for ensuring the safe implementation of machine-learning models in open environments. Although various algorithms for multi-class OOD detection have emerged, the exploration of multi-label OOD detection remains lacking. Current multi-label OOD algorithms ignore the inherent characteristic of label distribution imbalance in multi-label datasets, with limited attention paid to the scale imbalance issue between different objects in OOD detection. To address these challenges, we proposed a Multi-label Multi-scale Granularity-Aware OOD detection model (MMGA). Initially, a novel hierarchical multi-scale architecture was developed to extract critical information across various levels and granularities. This included the construction of a Scale-Aware Module (SAM) to align low-level details with high-level semantic features. Subsequently, to mitigate the negative impact of class-imbalanced data, a focal weighting strategy was introduced to exploit the potential of negative samples for OOD detection. Finally, we employed an energy function to calculate the OOD scores, thereby facilitating a reliable OOD uncertainty estimation. Experimental findings on the MS-COCO, PASCAL-VOC, and NUS-WIDE datasets confirm that our MMGA approach effectively addresses both imbalance issues, demonstrating superior performance compared to the JointEnergy method, with improvements of 7.62%, 9.52%, and 7.47% on the FPR95 metric, respectively.</p>