Abstract:
A key step for achieving the semantic segmentation of the test paper is to separate the printed and handwritten regions. In order to improve the effect of the semantic segmentation of the test paper, this paper proposes an improved attention algorithm based on the MaskRCNN network. By embedding the Subspace Multiscale Feature Fusion (SMFF) module into the feature pyramid structure of the MaskRCNN network, the attention features are calculated via the subspace such that the spatial and channel redundancy in the feature map can be reduced. By multi-scale feature fusion, the features of text regions with different sizes can be effectively extracted and the correlation between features can be enhanced. The experimental results show that for the target detection and semantic segmentation tasks of the test paper image dataset, the MaskRCNN network model based on the SMFF module can increase the average accuracy by 15.8% and 10.2% higher than that of the original MaskRCNN network model. Moreover, it also has greater performance improvement than the MaskRCNN based on the commonly used attention module.