作者:张培培，吕震宇
Authors:ZHANG Pei-pei,LU Zhen-yu摘要:摘要:随着深度学习框架的发展，新的目标检测算法也不断被提出，如一阶段、二阶段检测模型等，它们很好地提高了检测速度、解决了不同尺度目标检测的问题，但对于交叠、遮挡等问题，仍没能很好地解决。造成该问题的原因之一就在于模型训练期间，标签分配工作没有做好。针对该问题，提出基于全局信息的目标检测标签分配方法，该方法在模型训练阶段，利用指派方法，根据损失函数，建立全局最优的标签分配数学模型，给出了该模型与其他目标检测模型的融合方式，以及该方法在目标检测过程中所起到的作用。将该模型与常规的目标检测模型相融合，进行目标检测，实验结果表明，在有交叠、遮挡等复杂检测场景下，融合该方法的检测模型，其精确率均优于没有使用该方法的模型。
Abstract:Abstract:With the development of deep learning framework, new object detection algorithms have also been proposed, such as first-stage and two-stage detection models, which have improved the detection speed and solved the problem of object detection at different scales, but they have not yet been well solved for overlapping, occlusion and other issues. One of the reasons for this problem is that during model training, label assignment is not done well. Aiming at this problem, this paper proposes a target detection label allocation method based on global information, which uses the assignment method to establish a global optimal label allocation mathematical model based on the loss function in the model training stage, and gives the fusion mode of the model with other object detection models, and the role played by the method in the process of object detection. The experimental results show that the detection model of the fusion method is better than that of the model that does not use the method under the complex detection scenarios such as overlapping and occlusion.

PDF全文下载地址:

可免费Download/下载PDF全文