Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

How Panel Layouts Define Manga: Insights from Visual Ablation Experiments

Siyuan Feng, Teruya Yoshinaga, Katsuhiko Hayashi, Koki Washio, Hidetaka Kamigaito

TL;DR

The paper addresses how manga panel layouts contribute to the unique identity of a work. It leverages a visual ablation approach on facing-page manga images from the Manga109 dataset, creating original, text-and-character-masked, and panel-frame-only inputs, and trains a ResNet101-based classifier to predict titles; Grad-CAM visualizations accompany the results to interpret layout-focused decisions. The findings show that panel layouts alone carry strong work-specific information, achieving high accuracies across input types (e.g., 87.5% unprocessed and 84.3% panel-frame-only for 104-title classification), while publishers and genres exert only weak influence. This suggests that layout design is a defining feature of manga identity and highlights the potential for layout-centric analyses to illuminate authors’ stylistic choices. The work motivates further research with larger datasets and deeper analysis of intra-author page design differences to better understand how layout and content interplay in shaping manga style.

Abstract

Today, manga has gained worldwide popularity. However, the question of how various elements of manga, such as characters, text, and panel layouts, reflect the uniqueness of a particular work, or even define it, remains an unexplored area. In this paper, we aim to quantitatively and qualitatively analyze the visual characteristics of manga works, with a particular focus on panel layout features. As a research method, we used facing page images of manga as input to train a deep learning model for predicting manga titles, examining classification accuracy to quantitatively analyze these features. Specifically, we conducted ablation studies by limiting page image information to panel frames to analyze the characteristics of panel layouts. Through a series of quantitative experiments using all 104 works, 12 genres, and 10,122 facing page images from the Manga109 dataset, as well as qualitative analysis using Grad-CAM, our study demonstrates that the uniqueness of manga works is strongly reflected in their panel layouts.

How Panel Layouts Define Manga: Insights from Visual Ablation Experiments

TL;DR

The paper addresses how manga panel layouts contribute to the unique identity of a work. It leverages a visual ablation approach on facing-page manga images from the Manga109 dataset, creating original, text-and-character-masked, and panel-frame-only inputs, and trains a ResNet101-based classifier to predict titles; Grad-CAM visualizations accompany the results to interpret layout-focused decisions. The findings show that panel layouts alone carry strong work-specific information, achieving high accuracies across input types (e.g., 87.5% unprocessed and 84.3% panel-frame-only for 104-title classification), while publishers and genres exert only weak influence. This suggests that layout design is a defining feature of manga identity and highlights the potential for layout-centric analyses to illuminate authors’ stylistic choices. The work motivates further research with larger datasets and deeper analysis of intra-author page design differences to better understand how layout and content interplay in shaping manga style.

Abstract

Today, manga has gained worldwide popularity. However, the question of how various elements of manga, such as characters, text, and panel layouts, reflect the uniqueness of a particular work, or even define it, remains an unexplored area. In this paper, we aim to quantitatively and qualitatively analyze the visual characteristics of manga works, with a particular focus on panel layout features. As a research method, we used facing page images of manga as input to train a deep learning model for predicting manga titles, examining classification accuracy to quantitatively analyze these features. Specifically, we conducted ablation studies by limiting page image information to panel frames to analyze the characteristics of panel layouts. Through a series of quantitative experiments using all 104 works, 12 genres, and 10,122 facing page images from the Manga109 dataset, as well as qualitative analysis using Grad-CAM, our study demonstrates that the uniqueness of manga works is strongly reflected in their panel layouts.
Paper Structure (9 sections, 8 figures, 4 tables)

This paper contains 9 sections, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Experiment
  3. Multi-class Classification for 104 Works
  4. Qualitative Analysis of Classification Results
  5. Publisher Classification and Genre Classification
  6. Classification Experiments with Noisy Panel Frame-only Facing Page Images
  7. Rectangular noisy panels (Figure \ref{['fig:noise_rectangle']}):
  8. Quadrilateral noisy panels (Figure \ref{['fig:noise_quadrilateral']}):
  9. Conclusion

Figures (8)

  • Figure 1: Brief description of the multi-class manga title classification experiment of the titles of the comics. The images shown in the figure are from AisazuNihalrarenai © Masako Yoshi.
  • Figure 2: Examples of images used in the experiment. The images shown in the figure are from LoveHina © Ken Akamatsu.
  • Figure 3: Mean training and validation accuracy curves during model training, with standard deviation represented by error bars.
  • Figure 4: Grad-CAM heatmaps with unprocessed images. The model correctly classified input images and exhibited strong feature focus on areas where characters were drawn.
  • Figure 5: Grad-CAM heatmaps for page 49 of LoveHina_vol14 © Ken Akamatsu, with text-and-character-masked facing page images and panel frame-only facing page images.
  • ...and 3 more figures