This model weights file is designed to detect text in natural images. Unlike older methods that treated text as a single bounding box, CRAFT uses a character-level awareness approach. It generates a score map predicting the center of each character and an affinity map predicting the space between characters. This allows it to effectively detect arbitrary shapes (curved, rotated, or skewed text).
: CRAFT works by predicting "character region scores" and "affinity scores" (the likelihood that two characters belong to the same word). This allows it to handle curved or irregular text better than standard rectangular detectors.
: Unlike models that recognize letters (OCR), CRAFT focuses on detecting text regions. It identifies bounding boxes where text exists so a recognition model can later read it.
The name craft_mlt_25k.pth is explicitly derived from its training dataset and target metrics:
[Input Image] ──> [VGG-16 Backbone + U-Net Decoder] │ ├──> Channel 1: Region Score Map (Character Locations) └──> Channel 2: Affinity Score Map (Grouping Links) Dataset and Model Parameters