ScanSSD-XYc: Faster Detection for Math Formulas
ScanSSD-XYc: Faster Detection for Math Formulas
复制标题
ScanSSD-XYc：更快地检测数学公式
DOI：
10.1007/978-3-030-86198-8_7
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Zanibbi, Richard
中科院分区：
文献类型：
--
作者：
Dey, Abhisek;Zanibbi, Richard
Detecting formulas offset from text and embedded within textlines is a key step for OCR in scientific documents. The Scanning Single Shot Detector (ScanSSD) detects formulas using visual features, applying a convolutional neural network within windows in a large document image (600 dpi). Detections are pooled to produce page-level bounding boxes. The system works well, but is rather slow. In this work we accelerate ScanSSD in multiple ways: (1) input and output routines have been replaced by matrix operations, (2) the detection window stride (offset) can now be adjusted separately for training and testing, with fewer windows used in testing, and (3) merging with non-maximal suppression (NMS) in windows and pages has been replaced by merging overlapping detections using XY-cutting at the page level. Our fastest model processes 3 pages per second on a Linux system with a GTX 1080Ti GPU, Intel i7-7700K CPU, and 32 GB of RAM.
