Skip to content
This repository has been archived by the owner on Feb 11, 2024. It is now read-only.

Latest commit

 

History

History
68 lines (53 loc) · 2.34 KB

le2019deep.md

File metadata and controls

68 lines (53 loc) · 2.34 KB
Raw

Deep Learning Approach for Receipt Recognition

Le Duc, Anh & Pham, Dung & Nguyen, Tuan

Browse

@article{le2019deep,
   title={Deep Learning Approach for Receipt Recognition},
   ISBN={9783030356538},
   ISSN={1611-3349},
   url={http://dx.doi.org/10.1007/978-3-030-35653-8_50},
   DOI={10.1007/978-3-030-35653-8_50},
   journal={Lecture Notes in Computer Science},
   publisher={Springer International Publishing},
   author={Le, Anh Duc and Pham, Dung Van and Nguyen, Tuan Anh},
   year={2019},
   pages={705–712}
}

Pipeline

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

Receipt localization

  • We propose a simple method by the histograms of the input image on the x-axis and y-axis.

    (1) We calculate the histogram of the input image on x- axis and y-axis.

    (2) We determine the largest receipt area on each axis when the histogram is larger than a threshold. Then, we determine the receipt area.

Text line segmentation

  • The original CTPN detects horizontally arranged text. The CTPN structure is basically similar to Faster R-CNN, but with the addition of the LSTM layer.

  • (1) VGG16 extract feature from an input receipt. The size of the features is W×H×C.

    (2) A sliding window on the output feature map is employed. In each row, the sequential windows connected to a Bi-directional LSTM (BLSTM).

    (3) The BLSTM layer is connected to a 512D fully- connected layer. Then, the output layer predicts text/non-text scores, y-axis coordinates, and side-refinement offsets of 10.

Optical character recognition

  • The AED has two main modules: DenseNet for extracting features from a text image and an LSTM combined with an attention model for predicting the output text.

  • The context vector is computed by the attention mechanism.

Notes