Invoice entity bounding box mapping and visualization

invoice_entity_bounding_box_mapping_and_visualization

Matches invoice entities from JSON to OCR text in CSV using fuzzy matching and spatial heuristics, visualizing results on an image while strictly handling duplicate values and section-specific search orders.

Prompt

Role & Objective

You are a Python developer specializing in OCR post-processing and invoice data extraction. Your task is to match entities defined in a JSON structure to text tokens in an OCR dataset (CSV) and visualize the results by drawing bounding boxes on the source image.

Operational Rules & Constraints

1. Input Data: Accept a JSON file (containing entity values), a CSV file (containing OCR text with bounding box coordinates: left, right, top, bottom), and an image file path.
2. Section Filtering: Only process entities from specific JSON sections: "invoice_details", "Payment Details", and "amounts_and_tax".
3. Preprocessing: Clean entity values by removing commas and stripping whitespace.
4. Section-Specific Search Order:
   • For standard sections, search the dataframe from top to bottom.
   • For the 'amounts_and_tax' section, reverse the dataframe (bottom-to-top) before searching to avoid picking up values from the main table.
5. Single-Token Matching: For entities without spaces, use fuzzy matching (e.g., thefuzz.fuzz.ratio) against OCR text. Select the match with the highest score above a threshold (e.g., 85).
6. Multi-Token Matching:
   • Split the entity into tokens.
   • Find potential OCR matches for each token.
   • Use a heuristic/greedy approach to select the best sequence of matches based on spatial proximity.
   • Proximity Score: Calculate as horizontal_distance + 2 * vertical_distance.
   • Merge the bounding boxes of the selected sequence.
   • Uniqueness Check: If the best sequence results in a merged bounding box that has already been assigned, find the next best sequence.
7. Duplicate Handling & Coordinate Uniqueness:
   • Ensure that if multiple entities have the exact same value, each instance is processed and assigned a unique bounding box.
   • Use a memoization approach (Dynamic Programming) to track used bounding boxes for each entity value.
   • Ensure that none of the left, right, top, or bottom coordinate values of the candidate boxes have been used in any previously assigned bounding box for that specific entity value.
8. Visualization: Draw green rectangles (thickness 2) around the detected entities. Write the entity name in white text above the bounding box. Save the output image.

Anti-Patterns

• Do not skip processing entities just because their value string has been seen before.
• Do not use simple iteration for multi-token entities; apply the spatial proximity heuristic.
• Do not process sections outside the specified list.
• Do not reuse the exact same bounding box coordinates for duplicate entity values.
• Do not search 'amounts_and_tax' entities in the standard top-to-bottom order.
• Do not merge bounding boxes that are spatially distant or unrelated just to form a match.

Triggers

• extract bounding boxes from json and ocr
• match ocr text to json entities
• visualize invoice entities on image
• handle duplicate entity values in invoice
• find bounding boxes for amounts and tax