Image to Text

A deep learning-based approach to learning the image-to-text conversion, built on top of the OpenNMT system. It is completely data-driven, hence can be used for a variety of image-to-text problems, such as image captioning, optical character recognition and LaTeX decompilation.

Take LaTeX decompilation as an example, given a formula image:

The goal is to infer the LaTeX source that can be compiled to such an image:

 d s _ { 1 1 } ^ { 2 } = d x ^ { + } d x ^ { - } + l _ { p } ^ { 9 } \frac { p _ { - } } { r ^ { 7 } } \delta ( x ^ { - } ) d x ^ { - } d x ^ { - } + d x _ { 1 } ^ { 2 } + \; \cdots \; + d x _ { 9 } ^ { 2 } 

The paper [What You Get Is What You See: A Visual Markup Decompiler] provides more technical details of this model.

Dependencies

  • torchvision: conda install torchvision

  • Pillow: pip install Pillow

Quick Start

To get started, we provide a toy Math-to-LaTex example. We assume that the working directory is OpenNMT-py throughout this document.

Im2Text consists of four commands:

  1. Download the data.

wget -O data/im2text.tgz http://lstm.seas.harvard.edu/latex/im2text_small.tgz; tar zxf data/im2text.tgz -C data/
  1. Preprocess the data.

onmt_preprocess -data_type img \
                -src_dir data/im2text/images/ \
                -train_src data/im2text/src-train.txt \
                -train_tgt data/im2text/tgt-train.txt -valid_src data/im2text/src-val.txt \
                -valid_tgt data/im2text/tgt-val.txt -save_data data/im2text/demo \
                -tgt_seq_length 150 \
                -tgt_words_min_frequency 2 \
                -shard_size 500 \
                -image_channel_size 1
  1. Train the model.

onmt_train -model_type img \
           -data data/im2text/demo \
           -save_model demo-model \
           -gpu_ranks 0 \
           -batch_size 20 \
           -max_grad_norm 20 \
           -learning_rate 0.1 \
           -word_vec_size 80 \
           -encoder_type brnn \
           -image_channel_size 1
  1. Translate the images.

onmt_translate -data_type img \
               -model demo-model_acc_x_ppl_x_e13.pt \
               -src_dir data/im2text/images \
               -src data/im2text/src-test.txt \
               -output pred.txt \
               -max_length 150 \
               -beam_size 5 \
               -gpu 0 \
               -verbose

The above dataset is sampled from the im2latex-100k-dataset. We provide a trained model [link] on this dataset.

Options

  • -src_dir: The directory containing the images.

  • -train_tgt: The file storing the tokenized labels, one label per line. It shall look like:

<label0_token0> <label0_token1> ... <label0_tokenN0>
<label1_token0> <label1_token1> ... <label1_tokenN1>
<label2_token0> <label2_token1> ... <label2_tokenN2>
...
  • -train_src: The file storing the paths of the images (relative to src_dir).

<image0_path>
<image1_path>
<image2_path>
...