Getting Started

Welcome to the official technical documentation for Manga OCR Mobile.

Usage

Image Preprocessing

Before inference, images must be preprocessed to match the model's input expectations:

1. Canvas Setup: Create a 224x224 canvas filled with white background.
2. Scaling: Scale the input image to fit within the 224x224 box while preserving its aspect ratio.
3. Centering: Center the scaled image on the canvas.
4. Normalization:
   • Convert pixel values to floats in the range [0, 1] (divide by 255).
   • Arrange data in NCHW format (Channels, Height, Width) with RGB channel order.
     • Conversion to grayscale and back to RGB may be needed for optimal performance on colored pages.
   • The final input tensor shape is [1, 3, 224, 224].

Benchmarks

Performance metrics for the FP16 model (INT8 dynamic is about 1.5x faster):

• Galaxy S24 (CPU)
  • Encoder: 15-25ms
  • Decoder: 3-6ms per token
• Galaxy A6 (2018)
  • Encoder: 150-250ms
  • Decoder: 35-45ms per token