DynaOOD-Net: Dynamic Progressive Feature Fusion and Energy Balancing for Robust Out-of-Distribution Detection

In the field of object detection in open environments, models are required to simultaneously recognize predefined categories and accurately distinguish unknown categories from non-object regions, all while addressing the challenges of imbalanced data distributions and multi-scale targets. To meet these demands, this paper proposes DynaOOD-Net, a novel detection framework designed to enhance model generalization performance through dynamic feature integration and energy-balanced regularization strategies. DynaOOD-Net consists of two core components: the Dynamic Progressive Feature Pyramid Network (DP-FPN) and the Balanced Energy Constraint Loss (BECL). DP-FPN incorporates a Multi-Branch Dynamic Fusion Block (MBDF) to enable dynamic feature interactions and progressive optimization, ensuring robust feature representation across complex scenarios. Meanwhile, BECL dynamically adjusts category constraints, effectively addressing the indistinct energy boundary between known and unknown objects and the class imbalance problem, thus improving the model’s ability to distinguish between in-distribution (ID) and out-of-distribution (OOD) samples. Additionally, Dynamic Progressive Feature Pyramid Network-Lite (DP-FPN-Lite) is proposed, which reduces computational and storage overhead by simplifying the network architecture while preserving the core concepts of DP-FPN. Experimental results on seven datasets demonstrate that DynaOOD-Net significantly improves the mAP for known categories and consistently outperforms existing state-of-the-art methods in unknown object detection metrics. These findings establish DynaOOD-Net as a robust and efficient solution to the challenges inherent in open-set object detection. At the same time, DynaOOD-Lite utilizes DP-FPN-Lite to retain strong OOD detection capabilities while significantly reducing model parameters, making it suitable for deployment in resource-constrained environments.