This repository provides tools for quality assessment of digitalized documents . The estimated quality scores closely correnspond to readability by humans. The tools provide quality score heatmaps and an overall quality score for a whole document page. The module computes local perceptual quality scores based on confidence scores from Optical Character Recognition (OCR). More detailed description is provided below.
This module is build on top of OCR developed in project PERO (pero-ocr). The text recognition works in multiple stages. Firstly, locations and heights of text lines are determined using a fully convolutional neural network (modified U-NET). The individual text lines are processed by covolutional-recurrent networks trained using CTC loss. These networks provide confidences of recognized characters which are locally mapped to perceptual scores. The mapping to perceptual scores was calibrated on a large dataset of readability ratings by human readers.
This module requires python 3. Using virtualenv is highly recommended.
git clone --recursive https://github.com/DCGM/pero-quality.git
cd pero-quality
pip install -r requirements.txtNote: One of the requirements is tensorflow-gpu. If you don't have CUDA capable GPU installed in your machine, you can instead use tensorflow module to run the computation on CPU. If you have a CUDA capable GPU, but don't have CUDA drivers installed , refer to instructions for tensorflow GPU support.
Before processing a document, you need to download configuration and models for line detection and character recongnition. This is done by launching models/download_models.py script, or using the link located below.
All quality assessment functionality is encapsulated in QualityEvaluator class located in quality_evaluator module. Working with the module is shown below:
import cv2
from quality_evaluator import QualityEvaluator
quality = QualityEvaluator(config_path)
image = cv2.imread(filepath, cv2.IMREAD_COLOR)
score, heatmap = quality.evaluate_image(image)The output score holds scalar quality value of the whole page and heatmap is a numpy array of local quality at the input image resolution. More verbose example is located in pero_quality module.
Note: pero and pero_quality modules must be in PYTHONPATH to use QualityEvaluator class:
export PYTHONPATH=/abs/path/pero-quality/pero:/abs/path/pero-quality/pero_quality:$PYTHONPATH
To process multiple images from command line and render local quality overlays, launch the module as:
python3 pero-quality.py config/path input/path -o output/path -vBoth input path and output path can be directories, in which case all images in the input directory are processed. Alternatively single input and output file can be specified. Additional info about parameters is printed using single -h argument.
| Lidove noviny_2019-12-16 | This model is optimized for low-quality scans of czech newspapers scanned from microfilms. E.g. Lidove noviny. |
|---|---|
| Lidove noviny regression 2020-21-1 | Simplified faster model trained on low-quality scans of czech newspapers scanned from microfilms. E.g. Lidove noviny. |
| more will be added |
If you don't want to install this package and all dependencies on your system, you can use a provided dockerfile, which will create a separate image and docker container for this module.
docker directory contains basic info about how to run this repository in docker. We provide a dockerfile, and scripts for building an image and running docker container.
-
nvidia docker support
Note: use sudo with docker commands if user is not in docker group
-
build image:
build.sh -
create and run container:
run.sh -
run
pero-quality.pyor useQualityEvaluatoras needed