Meiki Text Recognition v0

meiki.txt.recognition.v0 is an image-to-text model that performs character detection for Japanese text, optimized for video game content. It redefines text recognition as a character detection task using a fine-tuned D-FINE object detector with a MobileNetV4 CNN backbone, outputting individual characters with bounding boxes and confidence scores.

Strengths

The model is designed to be pareto-optimal for Japanese text recognition, offering a superior accuracy-to-latency tradeoff compared to other open-weight alternatives. It is specifically trained on Japanese video games, where it significantly outperforms general-purpose OCR tools such as PaddleOCR or EasyOCR. The architecture is efficient on both CPU and GPU, with the companion detection model achieving typical latencies of ~30 ms (CPU) / ~3 ms (GPU) for the tiny variant.

Benchmarks

The following charts illustrate accuracy versus latency on CPU and GPU for the recognition model:

CPU	GPU

Input and output

Input images must be resized and padded to 960x32 pixels. The model detects horizontal text only, with a maximum of 48 characters per inference. Confidence threshold defaults to 0.1 in the inference script, while a value of 0.2 is recommended in the full pipeline CLI. Output is a JSON structure with each character, its bounding box coordinates, and a confidence value. The model is released under LGPL-3.0 (license discrepancy noted with the parent repository's Apache 2.0 badge).