基于双层解耦策略和注意力机制的遮挡目标分割

Occluded Object Segmentation Based on Bilayer Decoupling Strategy and Attention Mechanism

遮挡目标分割是实例分割中的一个难点,但在多个应用领域有很强的实用价值,例如物流传输线上堆叠快递包裹的分割。针对快递包裹目标遮挡导致难以分割的问题,该文提出一种基于双层解耦策略和注意力机制的遮挡目标分割方法。该方法首先利用带有特征金字塔(FPN)的主干网络提取图像特征;然后,利用双层解耦检测头自动预测实例的重心是否被遮挡并使用不同的分支对两类不同遮挡类型的实例进行检测;接下来,利用注意力改进模块得到无遮挡实例的预测掩模并将这些掩模合成为一个注意力权重图;最后,注意力改进模块利用该注意力权重图帮助有遮挡实例得到分割结果。该研究采集了一个遮挡快递包裹实例分割数据集,并在该数据集上进行实验。实验结果表明,该方法的平均精度(AP)、召回率(Recall)和漏检率(MR–2)指标分别达到了95.66%, 97.17%和11.78%,较其他方法具有更优的分割性能。

Occluded object segmentation is a difficult problem in instance segmentation, but it has great practical value in many industrial applications such as stacked parcel segmentation on logistics automatic sorting. In this paper, an occluded object segmentation method based on bilayer decoupling strategy and attention mechanism is proposed to improve the segmentation performance of occluded parcels. Firstly, the image features are extracted through a backbone network with a Feature Pyramid Network(FPN);Secondly,the bilayer decoupling head is used to predict whether the mass centers of instances are occluded, and different occlusion types of instances are predicted through different branches;Thirdly, attention refinement module is used to obtain predicted masks of non-occluded instances and generate an attention map by combining these masks;Finally, this attention map is used to help the prediction of occluded instances. A dataset is provided for occluded parcel segmentation. Our method is tested on this dataset. The experimental results show that the proposed network achieves 95,66% Average Precision(AP), 97.17% Recall, and 11.78% Miss Rate(MR–2). It indicates that this method has better segmentation performance than other methods.