CAI Meza is a software system that receives cell phone photos of log books, digitizes their contents and inserts them into a database. Meza ‘reads’ the hand-written log-book through optical character recognition into a digitized database. Data can then be accessed by various stakeholders in the format they need for reporting and analysis. CAI Meza offers an affordable, reliable, easy and impactful solution so that remote, low-tech environments are able to get the critical data all stakeholders need at the lowest cost.
Meza Promo Video - A video explanation of the CAI Meza platform
CAI Meza currently supports Western Arabic Numerals, Eastern Arabic Numerals and Bubbles. For numerals, our neural network architectures use a CRNN-like model built in PyTorch. The models use an EfficientNetB0 as backbone and feature extractor, which feeds a 512-dimension latent vector to a bidirectional LSTM. For bubbles, our models splits the cell image to isolate each bubble and then uses a modified ResNet18 to detect if it's checked or not.
This repository contains the datasets we use to benchmark the accuracy of our system. It also contains scripts to benchmark the accuracy of other OCR providers such as,
CAI helps organizations understand their data through concrete insights. We offer unique and customized solutions to empower leaders doing social good. We focus on tackling pressing real-world issues that have the biggest impact, with partners such as the United Nations World Food Program and the National Democratic Institute. Our core belief is that data technology can and should be a key tool in the mission to make the world a better place.
The test data is composed of actual pictures of logbooks from our pilots in the Republic of Congo (except for estern arabic numerals, currently synthetic).
Images take two forms. In test_data/full_images/, the logbook images are in their raw format. In test_data/cell_images/, we cut the individual cells from the source images.
The test_data/full_images/ directory also has a json file containing the shape of each potential template as Meza is configured to match incoming images with preloaded templates. As an aside, preloading templates is a feature that allows Meza to organize images and reference each extracted cell. It also allows it to automatically rotate the image if it's not upright.
This dataset is used to test the ability to take a raw image and cut it into its individual cells. It also comes with a templates.json file where the structure of the logbook is defined. The cell values can be found in the values.json file.
Directory = ./test_data/full_images/
Dataset Size = 6 images
Format : RGB images of varying width and height
Compression : JPEG
The images contain values expressed by this regex,
^[-]{0,1}[0-9]*[.,]{0,1}[0-9]*$
Dataset Size = 4451 cell images
Format : RGB images of varying width and height
Compression : JPEG
We have not yet collected field images containing the arabic script. Our handwritten dataset comes from here.
The images contain values expressed by this regex,
^[-]{0,1}[٠-٩]*[.,]{0,1}[٠-٩]*$
Dataset Size = 1001 cell images
Format : RGB images of varying width and height
Compression : JPEG
The images contain varying numbers of bubbles.
Dataset Size = 4642 cell images
Format : RGB images of varying width and height
Compression : JPEG
The run.ipynb jupyter notebook contains the scripts to benchmark the different OCR providers on our test data. Here are the accuracy scores.
Here, accuracy is defined as the number of cells successfully cut out
-
CAI Meza = 100%
-
Amazon =
-
Human inter-coder agreement =
-
CAI Meza = 86.94%
-
Google =
-
Amazon =
-
Azure =
-
Human inter-coder agreement =
-
CAI Meza = 82%
-
Google =
-
Amazon =
-
Azure =
-
Human inter-coder agreement =
-
CAI Meza = 96.8%
-
Google =
-
Amazon =
-
Azure =
- Jean-Romain Roy - Primary Developer of CAI Meza - jeanromainroy