DDSnet: A Deep Document Segmentation with Hybrid Blocks Architecture Network

Abstract

In recent years, the development of document segmentation technology is gaining more and more attention in the area of semantic segmentation, which plays an important role in the task of understanding the structure of documents. Although this demand applying deep learning approaches has undergone continuing advancement, the document segmentation systems still suffer from low accuracy rate. This paper presents a new high-performance document segmentation algorithm, namely Deep Document Segmentation Network (DDSnet), which incorporates advanced end-to-end deep learning methods for segmenting four different types of document features, including background, texts, tables, and figures. To overcome the small receptive fields, the atrous residual block is proposed, which is efficiently boosted by the adoption of multi-branches structure. For the better fine-grained output, the proposed atrous convolution residual block is conducted to achieve high accuracy. Moreover, this paper also releases the brand-new large-scale database, namely PPSD2019, for document segmentation that provides a pixel-level database for another benchmark. As documented in the experimental results, the proposed document segmentation method achieves a superior segmentation rate than that of the former competitive schemes. As a result, the proposed method and database can be considered as a very competitive candidate for the document segmentation applications.