Tesseract OCR Lambda Layer

AWS Lambda layer containing the tesseract OCR libraries and command-line binary for Lambda Runtimes running on Amazon Linux 1 and 2.

⚠️ The Amazon Linux AMI (Version 1) is being deprecated. Users are advised to not use Lambda runtimes (i.e. Python 3.6) based on this version. Refer also to the AWS Lambda runtime deprecation policy.

• Quickstart
• Ready-to-use binaries
  • Use with Serverless Framework
  • Use with AWS CDK
• Build tesseract layer from source using Docker
  • available Dockerfiles
  • Building a different tesseract version and/or language
  • Deployment size optimization
  • Building the layer binaries directly using CDK
  • Layer contents
• Known Issues
  • Avoiding Pillow library issues
  • Unable to import module 'handler': cannot import name '_imaging'
• Contributors ❤️

Quickstart

This repo comes with ready-to-use binaries compiled against the AWS Lambda Runtimes (based on Amazon Linux 1 and 2). Example Projects in Python 3.6 (& 3.8) using Serverless Framework and CDK are provided:

## Demo using Serverless Framework and prebuilt layer
cd example/serverless
npm ci
npx sls deploy

## or ..

## Demo using CDK and prebuilt layer
cd example/cdk
npm ci
npx cdk deploy

Ready-to-use binaries

For compiled, ready to use binaries that you can put in your layer see ready-to-use, or check out the latest release.

See examples for some ready-to-use examples.

Use with Serverless Framework

Serverless Framework

Reference the path to the ready-to-use layer contents in your serverless.yml:

service: tesseract-ocr-layer

provider:
  name: aws

# define layer
layers:
  tesseractAl2:
    # and path to contents
    path: ready-to-use/amazonlinux-2
    compatibleRuntimes:
      - python3.8

functions:
  tesseract-ocr:
    handler: ...
    runtime: python3.8
    # reference layer in function
    layers:
      - { Ref: TesseractAl2LambdaLayer }
    events:
      - http:
          path: ocr
          method: post

Deploy

npx sls deploy

Use with AWS CDK

AWS CDK

Reference the path to the layer contents in your constructs:

const app = new App();
const stack = new Stack(app, 'tesseract-lambda-ci');

const al2Layer = new lambda.LayerVersion(stack, 'al2-layer', {
    // reference the directory containing the ready-to-use layer
    code: Code.fromAsset(path.resolve(__dirname, './ready-to-use/amazonlinux-2')),
    description: 'AL2 Tesseract Layer',
});
new lambda.Function(stack, 'python38', {
    // reference the source code to your function
    code: lambda.Code.fromAsset(path.resolve(__dirname, 'lambda-handlers')),
    runtime: Runtime.PYTHON_3_8,
    // add tesseract layer to function
    layers: [al2Layer],
    memorySize: 512,
    timeout: Duration.seconds(30),
    handler: 'handler.main',
});

Build tesseract layer from source using Docker

You can build layer contents manually with the provided Dockerfiles.

Build layer using your preferred Dockerfile:

## build
docker build -t tesseract-lambda-layer -f [Dockerfile.al1|Dockerfile.al2] .
## run container
export CONTAINER=$(docker run -d tesseract-lambda-layer false)
## copy tesseract files from container to local folder layer
docker cp $CONTAINER:/opt/build-dist layer
## remove Docker container
docker rm $CONTAINER
unset CONTAINER

available Dockerfiles

Dockerfile	Base-Image	compatible Runtimes
Dockerfile.al1 (:warning: deprecated)	Amazon Linux 1	Python 2.7/3.6/3.7, Ruby 2.5, Java 8 (OpenJDK), Go 1.x, .NET Core 2.1
Dockerfile.al2	Amazon Linux 2	Python 3.8, Ruby 2.7, Java 8/11 (Coretto), .NET Core 3.1

Building a different tesseract version and/or language

Per default the build generates the tesseract 4.1.3 (amazonlinux-1) or 5.2.0 (amazonlinux-2) OCR libraries with the fast german, english and osd (orientation and script detection) data files included.

The build process can be modified using different build time arguments (defined as ARG in Dockerfile.al[1|2]), using the --build-arg option of docker build.

Build-Argument	description	available versions
TESSERACT_VERSION	the tesseract OCR engine	https://github.com/tesseract-ocr/tesseract/releases
LEPTONICA_VERSION	fundamental image processing and analysis library	https://github.com/danbloomberg/leptonica/releases
OCR_LANG	Language to install (in addition to eng and osd)	https://github.com/tesseract-ocr/tessdata (<lang>.traineddata)
TESSERACT_DATA_SUFFIX	Trained LSTM models for tesseract. Can be empty (default), _best (best inference) and _fast (fast inference).	https://github.com/tesseract-ocr/tessdata, https://github.com/tesseract-ocr/tessdata_best, https://github.com/tesseract-ocr/tessdata_fast
TESSERACT_DATA_VERSION	Version of the trained LSTM models for tesseract. (currently - in July 2022 - only 4.1.0 is available)	https://github.com/tesseract-ocr/tessdata/releases/tag/4.1.0

Example of custom build

## Build a Dockerimage based on Amazon Linux 2, with French language support
docker build --build-arg OCR_LANG=fra -t tesseract-lambda-layer-french -f Dockerfile.al2 .
## Build a Dockerimage based on Amazon Linux 2, with Tesseract 4.0.0 and french language support
docker build --build-arg TESSERACT_VERSION=4.0.0 --build-arg OCR_LANG=fra -t tesseract-lambda-layer -f Dockerfile.al2 .

Deployment size optimization

The library files that are content of the layer are stripped, before deployment to make them more suitable for the lambda environment. See Dockerfiles:

RUN ... \
  find ${DIST}/lib -name '*.so*' | xargs strip -s

The stripping can cause issues, when the build runtime and the lambda runtime are different (e.g. if building on Amazon Linux 1 and running on Amazon Linux 2).

Building the layer binaries directly using CDK

You can build the layer directly and get the artifacts (like in ready-to-use). This is done using AWS CDK with the bundling option.

Refer to continous-integration and the corresponding Github Workflow for an example.

Layer contents

The layer contents get deployed to /opt, when used by a function. See here for details. See ready-to-use for layer contents for Amazon Linux 1 and Amazon Linux 2 (TODO).

Known Issues

Avoiding Pillow library issues

Use cloud9 IDE with AMI linux to deploy example. Or alternately follow instructions for getting correct binaries for lambda using EC2. AWS lambda uses AMI linux distro which needs correct python binaries. This step is not needed for deploying layer function. Layer function and example function are separately deployed.

Unable to import module 'handler': cannot import name '_imaging'

You might run into an issue like this:

/var/task/PIL/_imaging.cpython-36m-x86_64-linux-gnu.so: ELF load command address/offset not properly aligned
Unable to import module 'handler': cannot import name '_imaging'

The root cause is a faulty stripping of libraries using strip here.

Quickfix

You can just disable stripping (comment out the line in the Dockerfile) and the libraries (*.so) won't be stripped. This also means the library files will be larger and your artifact might exceed lambda limits.

A lenghtier fix

AWS Lambda Runtimes work on top of Amazon Linux. Depending on the Runtime AWS Lambda uses Amazon Linux Version 1 or Version 2 under the hood. For example the Python 3.8 Runtime uses Amazon Linux 2, whereas Python <= 3.7 uses version 1.

The current Dockerfile runs on top of Amazon Linux Version 1. So artifacts for runtimes running version 2 will throw the above error. You can try and use a base Dockerimage for Amazon Linux 2 in these cases:

FROM: lambci/lambda-base-2:build
...

or, as @secretshardul suggested

simple solution: Use AWS cloud9 to deploy example folder. Layer can be deployed from anywhere. complex solution: Deploy EC2 instance with AMI linux and get correct binaries.

Contributors ❤️

• @secretshardul
• @TheLucasMoore for providing a Dockerfile that builds working binaries for Python 3.8 / Amazon Linux 2