摘要： 为避免锚点对目标检测模型的鲁棒性造成负面影响，并保证在无锚点情况下模型具有较高的准确度，提出一种单线程无锚点全卷积网络模型。通过取消预设锚点参数以及像素级别预测，使得模型在无锚点情况下检测目标时具有更高的鲁棒性。使用沙漏骨干网络取代特征金字塔模块，从而降低锚点与特征金字塔模块的冗余以及计算量，使整体模型结构更加精简。实验结果表明，与经典锚点Retinanet模型相比，该模型利用正例区域原则与中心偏离支路显著提高了预测能力，并且具有更高的正负标签比例和更快的推理速度。

关键词: 目标检测, 无锚点网络, 单线程, 模型鲁棒性, 像素级预测

Abstract: In order to avoid the negative impact of anchor on the robustness of the target detection model and ensure the model has high accuracy without anchor, this paper proposes a single thread and no anchor full convolution network model. By canceling the pre-set anchor parameters and pixel level prediction, the model has higher robustness when detecting targets without anchor. The hourglass backbone network is used to replace the feature pyramid module, which reduces the redundancy and calculation of anchor and feature pyramid module, and makes the overall model structure more simplified. The experimental results show that compared with the classical Retinanet model, the proposed model can significantly improve the prediction ability by using the positive case region principle and the center deviation branch, and has higher positive and negative label ratio and faster reasoning speed.

Key words: object detection, anchor-free network, single route, robustness of the model, pixel level prediction

中图分类号: 

• TP18