基于多尺度四维特征融合的小样本语义分割

Few-shot semantic segmentation based on multiscale 4D feature fusion

现有的语义分割方法依赖充足的像素级图像标签,而分割模型需要在新类样本条件下进行训练时,带来人工标注图像的问题;因此,提出了小样本语义分割方法来解决此类问题.当前小样本分割方法主要采用原型学习方法,而原型学习的方法缺乏像素级支持特征来指导查询图像分割,导致分割精度不高的问题.基于此,设计了一种四维特征融合与注意力增强的小样本分割网络.为了获取到像素级支持特征对查询图像的表征信息,在特征金字塔结构中使用四维卷积,将高级语义特征和中级语义特征逐步压缩成超相关特征进而应用于查询图像的分割中.同时,在两个标准小样本分割基准上进行了实验:在PASCAL-5i数据集1-shot设置下的测试结果mIoU分别比HSNet和PFENet提高了0.6%和2.3%.

The existing semantic segmentation method relies on sufficient pixel-level image labeling,while the segmentation model needs to be trained under the new sample conditions,which brings the problem of manually labeling images.Therefore,few-shot semantic segmentation method is proposed to solve such problems.The current few-shot segmentation method mainly adopts the prototype learning method,while the prototype learning method lacks pixel-level support-level features to guide query image segmentation,resulting in the problem of low segmentation accuracy.Based on this,a four-dimensional feature fusion and attention-enhancing few-shot segmentation network is designed.In order to obtain the pixel-level representation information of rich support set features for query images,four-dimensional convolution is used in the feature pyramid structure to gradually compress advanced semantic features and intermediate semantic features into super-correlated features,which is then used to segment the query image.At the same time,the test results of mIoU in the PASCAL-5i dataset 1-shot setting were improved by 0.6%and 2.3%compared to the HSNet and PFENet,respectively.