基于双阶段网络的交互式目标分割算法

Interactive Target Segmentation Algorithm Based on Two-Stage Network

针对已有多数交互式分割方法交互方式单一、预测结果精度较低的问题,构建一种基于双阶段网络的目标分割模型ScribNet,以实现更完整和精细的交互式目标分割。采用灵活涂画的交互方式,通过编码形成交互指导信息。设计骨架随机性仿真算法,实现大数据下的模拟交互操作。在传统分割模型中引入预测优化模块,形成双阶段网络结构,以充分利用交互指导信息。在COCO和PASCAL数据集上的实验结果表明,与DEXTR、GrabCut等方法相比,ScribNet模型的分割精度较高。

Most of the existing interaction segmentation methods are faced with the single interaction mode and inaccurate prediction results.To address the problem,this paper proposes a target segmentation model called ScribNet based on two-stage network to realize more complete and precise interactive target segmentation.The model adopts the scribbling interaction mode,and encodes the information to form interaction guidance.A skeleton randomness simulation algorithm is designed to perform simulated interactions under big data. Then a prediction optimization module is introduced into the traditional segmentation model to form a two-stage network architecture to fully use the interaction guidance information.Experimental results on COCO and PASCAL datasets show that compared with DEXTR,GrabCut and other methods,the proposed ScribNet has a higher segmentation accuracy.