dai2017fused.md

Fused Text Segmentation Networks for Multi-oriented Scene Text Detection

Yuchen Dai, Zheng Huang, Yuting Gao, Youxuan Xu, Kai Chen, Jie Guo, Weidong Qiu

Browse

@article{dai2017fused,
   title={Fused Text Segmentation Networks for Multi-oriented Scene Text Detection},
   ISBN={9781538637883},
   url={http://dx.doi.org/10.1109/ICPR.2018.8546066},
   DOI={10.1109/icpr.2018.8546066},
   journal={2018 24th International Conference on Pattern Recognition (ICPR)},
   publisher={IEEE},
   author={Dai, Yuchen and Huang, Zheng and Gao, Yuting and Xu, Youxuan and Chen, Kai and Guo, Jie and Qiu, Weidong},
   year={2018},
   month={Aug}
}

Pipeline

Receipt detection	Receipt localization	Receipt normalization	Text line segmentation	Optical character recognition	Semantic analysis
❌	❌	❌	✔️	❌	❌

Text line segmentation

• resnet-101 backbone

• PSROIPooling instead of ROIPooling + FC layers (faster inference time)

  

  source: https://arxiv.org/pdf/1908.03673.pdf

• source: https://arxiv.org/pdf/1908.03673.pdf

• source: http://deeplearning.csail.mit.edu/instance_ross.pptx

• source: https://joshua19881228.github.io/2016-05-23-RFCN/

• source: https://www.semanticscholar.org/paper/R-FCN%3A-Object-Detection-via-Region-based-Fully-Dai-Li/b724c3f7ff395235b62537203ddeb710f0eb27bb

• PSRoi Inside/Outside Pooling - more here

• smooth-L1

• Ground truth of each text instance is presented by bounding boxes and masks

• The post-processing part includ

• After NMS, we generate a minimum quadrilateral for each text instance covering the mask

• Standard NMS computes IOU among bounding boxes

• Mask-NMS mainly changes bounding box IOU computation to so-called mask-maximum-intersection (MMI)

• Maximum intersection over the mask areas are used to replace original IOU for the reason that detections may easily involve line-level and word-level text instances simultaneously at the same line

Notes

• 

• multi-oriented text detection from an instance aware segmentation perspective

• "Fused Text Segmentation Networks"

• Ground truth of each text instance is presented by bounding boxes and masks