Home /Modules/character-recognition/README.md

README.md

lib/character-recognition/README.md · 47 line(s) · markdown

SLeeLa Character Recognition

SLeeLa-facing package for the native C11 glyph-recognition subsystem in artificial-intelligence/character-recognition/. The native C API remains authoritative for normalization, sample storage, recognition, segmentation, model persistence, and transcript writing. The SLeeLa classes do not claim that those operations run in the VM without an installed host adapter.

Public units

Native bridge contract and limits

The adapter maps SLCharacterRecognition to character_recognition.h and the slcr_* C functions. Glyph buffers contain exactly 784 bytes in row-major 28×28 order. Grayscale input dimensions must match the input buffer; normalization emits the native binary convention (foreground 255, background 0). Native labels are one UTF-8 code point and fit in SLCR_LABEL_CAPACITY bytes including the terminator. The adapter must preserve the native model/buffer ownership rules and pair segment allocations with slcr_free_segments.

The native classifier recognizes clean, pre-segmented 28×28 glyphs using nearest-neighbour pixel similarity. The page segmenter is projection based and is intended for clean, separated writing. This is not a general OCR/layout engine: touching cursive letters, complex page layouts, deskew, language-model correction, and calibrated probability estimates are not provided.

A similarity score is not a probability or legal confidence rating. Preserve the original scan and review the transcript and JSONL sidecar. The native API does not modify the source image.

The adapter's core operations map to slcr_model_create, slcr_model_add, slcr_normalize_glyph, slcr_model_recognize, and slcr_model_destroy; persistence, segmentation, and transcript functions remain available through their documented native API.

Compile character_recognition.c as C11 and link the native object. PGM loading and segmentation allocate native buffers; ownership and matching free calls must be respected by the adapter.

Run native regression checks from the repository root:

make -C tests ai-character-recognition character-recognition-cli