面对主流计算平台对框架轻量化的需求，设计一种基于多任务结构的轻量化抠图框架。将总体任务拆分为两类子任务，其中一类任务用来在语义层面上为高级特征分类，区分前景背景与未知区域的特征;另一类任务用于计算前景与背景图层的线性组合权重。通过与特征分类任务共享高级特征网络的权值获得精准的前景特征，再与低级别卷积特征相融合。所提出的模型能够生成精准的抠图掩膜，同时优化卷积神经网络来实现模型轻量化。在Composition 1K数据集上对比不同方法的实验结果:在分辨率为640×640的输入条件下，所提方法比DIM(deep image matting)和AdaMatting(adaptation and matting)方法分别减少19%和81%的空间消耗;对于同样的数据输入，所提方法需要的处理时间只有DIM消耗时间的五分之一。
A lightweight image matting method based on attentive model for multi-level appearance cues
A lightweight image matting framework,which is based on multi-task structure,is designed to meet the requirements of the mainstream computing platforms.Concretely,the overall task can be split into two sub-tasks.One sub-task is to classify the higher-level features at the semantic level,and then it distinguishes foreground/background features from the unknown regions.Another task is to calculate the weights of the linear combination for the foreground and background layers.Accurate foreground features are obtained by sharing the weights of high-level feature networks with feature classification tasks,and they are fused with low-level convolution features.The proposed model outputs more accurate mattes.Also,the convolutional neural network is optimized to lightweight the model.On a benchmark dataset of Composition 1K,schemes performance is compared with different architectures.The proposal can reduce the 19%and 81%of storage consumption in comparison with DIM(deep image matting)and AdaMatting(adaptation and matting)on 640×640 images.For the identical data inputs,the running time of the proposed model is only about 1/5 of DIM's.
digital image mattinglightweighttrimapmulti-task frameworkmulti-level appearance cues
NETL
NSTL
万方数据
维普
