Artificial Intelligence/Robotoff/Adding new detections

From Open Food Facts wiki
Revision as of 14:07, 3 August 2020 by Teolemon (talk | contribs)
Jump to navigation Jump to search

Finding opportunities

  1. Download the OCR JSONL
  2. grep -o '\<WORD\>' | wc -l

Adding detections using REGEXes

https://github.com/openfoodfacts/robotoff/tree/master/robotoff/insights/ocr

Add a function to add the desired match

  • Write the function
def process_fsc_match(match) -> str:
    fsc_code = match.group(1)
    return "FSC-{}".format(fsc_code).upper()

Add the function to the processing pipeline to plug it to Robotoff

This happens in the same file.

PACKAGER_CODE: Dict[str, OCRRegex] = {
    "fr_emb": OCRRegex(
        re.compile(r"emb ?(\d ?\d ?\d ?\d ?\d) ?([a-z])?(?![a-z0-9])"),
        field=OCRField.text_annotations,
        lowercase=True,
        processing_func=process_fr_emb_match,
    ),
    "fsc": OCRRegex(
        re.compile(r"fsc.? ?(c\d{6})"),
        field=OCRField.text_annotations,
        lowercase=True,
        processing_func=process_fsc_match,
    ),
    "eu_fr": OCRRegex(
        re.compile(

Grab yourself a sandwich

Adding detections using Flashtext

https://github.com/openfoodfacts/robotoff/tree/master/data/ocr

Expanding existing detections

  • Choose a file matching the detection you want to add

In this example: https://github.com/openfoodfacts/robotoff/blob/master/data/ocr/brand_logo_annotation.txt

  • Add entries one per line. The raw `logoAnnotation` you find in the OCR JSON is on the left, the value you want to add in the brand field is on the right.
alnatura||Alnatura
Andros France||Andros
andros||Andros

Add a new type of detection

  • Create a file `fishing_flashtext.txt` and add one value to extract per line

Add it to the Robotoff settings

robotoff/settings.py
OCR_LABEL_WHITELIST_DATA_PATH = OCR_DATA_DIR / "label_whitelist.txt"
OCR_FISHING_FLASHTEXT_DATA_PATH = OCR_DATA_DIR / "fishing_flashtext.txt"

Plug it into the OCR pipeline

robotoff/insights/ocr/packager_code.py
    codes = text_file_iter(settings.OCR_FISHING_FLASHTEXT_DATA_PATH)
    return generate_keyword_processor(("{}||{}".format(c.upper(), c) for c in codes))