Semantic Segmentation of PHT Based on Improved DeeplabV3+

This work aimed to address the two shortcomings of printed and handwritten texts (PHT) classification. The classification accuracy FCN U-net, which are used for PHT pixel-level classification, still has room improve. public datasets have small sample sizes, generalization ability models is not good. In this paper, first, a sample-making method identification was proposed, dataset 2021 (PHTD 2021), containing 3,000 samples, constructed. Second, because there large number words but contours in documents, DeeplabV3+ model improved. network layer pooling times were reduced, convolution kernel dilated rate increased. experiment, improved had 95.06% on test samples from PHTD dataset. higher recognition than models. Finally, after PHT, applications removal extraction provided.