dm-codes / README.md
shortery's picture
add external links and citing to README.md
4f58d6d verified
metadata
dataset_info:
  features:
    - name: image
      dtype: image
    - name: tl
      sequence: int64
    - name: tr
      sequence: int64
    - name: br
      sequence: int64
    - name: bl
      sequence: int64
    - name: is_clean
      dtype: bool
    - name: split
      dtype: string
    - name: text
      dtype: string
    - name: image_name
      dtype: string
  splits:
    - name: validation
      num_bytes: 64018244
      num_examples: 101
    - name: test
      num_bytes: 125460818
      num_examples: 199
  download_size: 189448472
  dataset_size: 189479062
configs:
  - config_name: default
    data_files:
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: mit
task_categories:
  - image-to-text
  - object-detection
size_categories:
  - n<1K

DM codes dataset

The dataset contains photos of Data Matrix (DM) codes and their annotations. The photos were taken on an iPhone and annotated manually by a human. The annotations contain text, which is encoded in the DM code and the pixel coordinates of the DM code vertices. The vertices are: tl = top left, tr = top right, br = bottom right, bl = bottom left. Attribute is_clean specifies whether the DM code on the image is expected to be easily readable. For every DM code, there is exactly one image with is_clean=true and several images with is_clean=false.

If you want to crop the DM codes from the images, use the following code:

import numpy as np
import datasets
from PIL import Image
from skimage import transform

def crop_dm_code(example: dict, square_side: int = 200, square_padding: int = 25) -> dict:
    vertices = np.asarray((example["tl"], example["tr"], example["br"], example["bl"]))
    unit_square = np.asarray([
        [square_padding, square_padding],
        [square_side + square_padding, square_padding],
        [square_side + square_padding, square_side + square_padding],
        [square_padding, square_side + square_padding]
    ])
    transf = transform.ProjectiveTransform()
    if not transf.estimate(unit_square, vertices): raise Exception("estimate failed")
    cropped_np_image = transform.warp(
        np.array(example["image"]),
        transf,
        output_shape=(square_side + square_padding * 2, square_side + square_padding * 2)
    )
    cropped_image = Image.fromarray((cropped_np_image * 255).astype(np.uint8))
    return {"cropped_image": cropped_image}

dataset = datasets.load_dataset("shortery/dm-codes")
dataset = dataset.map(crop_dm_code)

DataMatrix Image Reconstruction to Enhance Decodability

This dataset is a part of the Diploma thesis https://is.muni.cz/th/ppu25/dp-dmcodes-thesis.pdf. This thesis compares various encoder-decoder CNNs to enhance the DM code image quality before decoding it with a code reader. The code is available on GitHub https://github.com/shortery/dp-dm-codes.

Citing

@thesis{dmcodes-thesis,
  author = {Petra Krátká},
  title = {DataMatrix Image Reconstruction to Enhance Decodability},
  address = {Brno},
  year = {2024},
  school = {Masaryk University, Faculty of Informatics},
  type = {Diploma thesis},
  url = {https://is.muni.cz/th/ppu25/dp-dmcodes-thesis.pdf},
}