Object detection is one of the most significant tasks in recent computer vision and healthcare study, which also has been applied in many areas. Although some detection frameworks show good performance for some specific datasets, the ambiguity in feature levels of anchor-free detectors still limits the performance of both fully-supervised and cross-dataset settings. Hence, a digital twins end-to-end multi-branch object detection framework with feature level selection is presented in this work. First, a five-level feature pyramid is adopted with a set of detection heads to construct an anchor-free detection backbone. Then, a learning-based selection strategy is presented to help obtain better feature level selection performance. Experimental results on general object detection datasets show that our framework can achieve 39.2 average precision (AP) on the COCO dataset and 10.2 miss rate (MR) on the CityPersons dataset. Furthermore, experimental results on cross-dataset settings, including Cityscapes, Caltech, SIM 10k, KITTI datasets, have also proved the good generalization ability of our framework. Through the optimized models in digital twins, it is also been applied in a pneumonia detection dataset with 49.3 AP. In addition, a large number of comparisons with state-of-the-art works also verify the detection performance and real-time efficiency of the proposed framework.

目标检测是最近计算机视觉和医疗保健研究中最重要的任务之一，它也已应用于许多领域。尽管某些检测框架在某些特定数据集上表现出良好的性能，但无锚检测器的特征级别的模糊性仍然限制了全监督和跨数据集设置的性能。因此，本文提出了一种具有特征级别选择的数字孪生端到端多分支对象检测框架。首先，采用带有一组检测头的五级特征金字塔来构建无锚检测主干。然后，提出了一种基于学习的选择策略，以帮助获得更好的特征级选择性能。在通用对象检测数据集上的实验结果表明，我们的框架可以达到 39。COCO 数据集的平均精度 (AP) 为 2，CityPersons 数据集的未命中率 (MR) 为 10.2。此外，跨数据集设置的实验结果，包括 Cityscapes、Caltech、SIM 10k、KITTI 数据集，也证明了我们框架的良好泛化能力。通过数字孪生中的优化模型，它也被应用在一个49.3 AP的肺炎检测数据集中。此外，与state-of-the-art作品的大量比较也验证了所提出框架的检测性能和实时效率。通过数字孪生中的优化模型，它也被应用在一个49.3 AP的肺炎检测数据集中。此外，与state-of-the-art作品的大量比较也验证了所提出框架的检测性能和实时效率。通过数字孪生中的优化模型，它也被应用在一个49.3 AP的肺炎检测数据集中。此外，与state-of-the-art作品的大量比较也验证了所提出框架的检测性能和实时效率。