Most deep learning-based object detection methods are proposed based on multi-level feature environments. Although some researchers have tried to detect on one-level features, where multiple feature maps are utilized. In this paper, we aim to propose a novel anchor-free object detection approach with an automatic label assignment strategy on only one feature map. The proposed method follows the main idea of AutoAssign to achieve the label assignment strategy. However, to make the strategy work appropriately in one feature map environment, several modifications have been made. A prior knowledge fusion method called ‘Center Weighting Fusion’ is proposed for label assignment strategy, where the Gaussian mixture model function is applied to calculate the weights of each object on one feature map. By doing so, some objects that are close to each other, whose weights will be merged and generate some points (Recheck Points) that are shared with multiple objects. For those ‘Recheck Points,’ the detector will judge how many objects share this point and generate a corresponding number of different-sized proposals. In detecting different-sized objects, we propose a ‘Uniform Detection’ method to limit each point’s regression distance according to the target’s category. A large number of experimental data show that the proposed method presents competitive detection accuracy with normal anchor-free detectors (43.8% mAP), while it is smaller (30% smaller) and faster (50% better).

大多数基于深度学习的目标检测方法都是基于多级特征环境提出的。尽管一些研究人员尝试检测使用多个特征图的一级特征。在本文中，我们的目标是提出一种新颖的无锚目标检测方法，仅在一个特征图上采用自动标签分配策略。该方法遵循自动分配的主要思想来实现标签分配策略。然而，为了使该策略在一个特征图环境中正常工作，进行了一些修改。针对标签分配策略，提出了一种称为“中心加权融合”的先验知识融合方法，其中应用高斯混合模型函数来计算一个特征图上每个对象的权重。通过这样做，一些彼此接近的对象的权重将被合并并生成一些与多个对象共享的点（重新检查点）。对于那些“重新检查点”，检测器将判断有多少对象共享该点并生成相应数量的不同大小的建议。在检测不同大小的物体时，我们提出了一种“统一检测”方法，根据目标的类别限制每个点的回归距离。大量实验数据表明，该方法与普通无锚检测器（mAP）具有竞争性的检测精度（43.8％），同时更小（小30％）和更快（更好50％）。