Poneglyph-Classifier

Poneglyph Page Type Classifier

A MobileNetV3-Small classifier trained on full manga volumes (One Piece tomes 1 to 7) to automatically detect page types:

• cover: Volume covers, inner covers, chapter cover/title pages
• story_page: Narrative manga reading pages
• annexe: SBS (question/answer corners), fan art galleries, author notes, bonus pages
• summary: Table of contents / volume summaries

---

Model Description

• Architecture: MobileNetV3-Small
• Input shape: [1, 3, 224, 224] (RGB)
• Output shape: [1, 4] (logits)
• Class order: ['cover', 'story_page', 'annexe', 'summary']
• Preprocessing:
  • Resize shortest edge to 256
  • Center crop 224x224
  • ImageNet normalization: Mean [0.485, 0.456, 0.406], Std [0.229, 0.224, 0.225]
• Formats provided:
  • page_type_classifier.onnx: Static FP32 ONNX model for browser (onnxruntime-web) and backend execution (onnxruntime)
  • final.pt: PyTorch weights checkpoint
  • page_type_classifier.metadata.json: Full preprocessing and class contract
  • metrics.json: Detailed training and evaluation metrics

---

Dataset Statistics

The dataset was directly extracted from authentic CBZ manga volumes (One Piece tomes 1 to 7) and 100% human-validated.

---

Evaluation Results (Held-out Tome 7)

Evaluation performed on One Piece Tome 7 (196 pages), completely held-out during training:

• Overall Accuracy: 99.49% (195 / 196 correct)
• Macro F1-Score: 0.9845
• Validation Loss: 0.0476

Detailed Per-Class Metrics

Confusion Matrix

               Predicted
               cover   story_page   annexe   summary
cover            8         1           0        0
story_page       0       164           0        0
annexe           0         0          22        0
summary          0         0           0        1

---

Usage

Python (ONNX Runtime)

import numpy as np
import onnxruntime as ort
from PIL import Image

session = ort.InferenceSession("page_type_classifier.onnx", providers=["CPUExecutionProvider"])
classes = ["cover", "story_page", "annexe", "summary"]

image = Image.open("page.jpg").convert("RGB")
w, h = image.size
scale = 256.0 / min(w, h)
image = image.resize((int(round(w * scale)), int(round(h * scale))), Image.Resampling.BILINEAR)

left = (image.width - 224) // 2
top = (image.height - 224) // 2
image = image.crop((left, top, left + 224, top + 224))

arr = np.array(image, dtype=np.float32) / 255.0
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
arr = (arr - mean) / std
arr = np.transpose(arr, (2, 0, 1))
tensor = np.expand_dims(arr, axis=0)

logits = session.run(None, {"input": tensor})[0][0]
probs = np.exp(logits - np.max(logits))
probs /= probs.sum()
predicted_class = classes[int(np.argmax(probs))]
print(predicted_class, probs)