SFNet: Stage Spatial Attention and Feature Waterfall Fusion for Mathematical Document Text Line Detection
Chang Liu, Jiaxin Liu, Yong Zhang, Lei Huang · 2021 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) · 2021
Text line detection is an important technique in layout analysis. Due to the existence of non-text areas and the change of formula height, text line detection in mathematical documents has always been a huge challenge. In view of the complexity of the spatial structure of mathematical text lines, on the one hand, based on the existing attention mechanism, we proposed a composite network structure with Stage Spatial Attention, which has better feature extraction. In addition, a multi-scale feature fusion structure, called Feature Waterfall Fusion, adds features with low-resolution but stronger semantic information into the features with high-resolution but weaker semantic information indirectly and directly. Combining them results in SFNet. On the other hand, to explore the mathematical document text line detection, we constructed a dataset named Mathematical Document Text Line Dataset (MDTL-1850), which includes about 10,669 text line annotations in 1,850 images (1,280 for training and 570 for testing). We assembled SFNet and a learnable post-processing Progressive Scale Expansion (PSE) surpassed mainstream text detection networks such as PSENet and PANet in all metrics on MDTL-1850. Precision$(P)$, Recall$(R)$and$F_{1}$, which improved by 1.0, 0.7 and 0.9 respectively. The experimental results strongly proved the superiority of SFNet network in detecting text line of mathematical documents.