Ablation CAM Analysis in CNNs

• The paper demonstrates that combining systematic ablation with Grad-CAM highlights the role of panel layouts, achieving up to 84.3% title accuracy with frame-only inputs.
• Ablation CAM employs precise visual ablation protocols to isolate semantic content, enabling analysis that distinguishes layout cues from characters and text.
• Integrating Grad-CAM yields spatial heatmaps that reveal distinctive layout features, underscoring the method's potential for enhancing interpretability in document analysis.

Ablation CAM is a framework for dissecting and visualizing the internal representations of convolutional neural networks (CNNs) through the combined use of systematic ablation (masking or isolating input features) and class activation mapping (CAM), particularly the Grad-CAM algorithm. In the context of manga analysis, Ablation CAM has been deployed to interrogate the discriminative power of panel layouts by isolating page structures from characters and text, enabling a rigorous evaluation of how layout alone can drive high-fidelity classification in deep learning pipelines (Feng et al., 2024).

1. Visual Ablation Protocols

Ablation procedures are explicitly designed to control the visibility of semantic content in input images. The canonical setup, as formalized by Feng et al., includes three distinct input types derived from the Manga109 dataset:

• Unprocessed images ($I_{\text{orig}}$): Raw, unmodified facing-page scans.
• Text/Character-masked ($I_{\text{masked}}$): Regions corresponding to characters and text are occluded by black rectangles, operationalized as $$I_{\text{masked}} = I_{\text{orig}} \odot (1 - M_{\text{char}})$$, where $M_{\text{char}}$ is a binary mask.
• Panel frame-only ($I_{\text{frame}}$): Only panel boundaries are rendered on a blank background, such that $I_{\text{frame}}(x, y) = 1$ where $(x, y)$ lies on any annotated panel edge; 0 otherwise.

The ablation is implemented via algorithmic masking and by drawing panel boundaries from dataset annotations, ensuring that semantic content besides panel structure is systematically excluded [(Feng et al., 2024), Sec. 3].

2. CNN Classifier and Experimental Design

A ResNet-101 architecture, initialized from ImageNet pretraining, constitutes the classification backbone, with its final layer adapted to a 104-way softmax to match the number of unique manga titles. The network receives 3×224×224 RGB images, resized from input, and is optimized via multi-class cross-entropy loss:

$$L = -\sum_{i=1}^N \sum_{c=1}^{104} y_{i,c} \log p_{i,c}$$

Stochastic gradient descent (SGD) with momentum 0.90, an initial learning rate of $10^{-3}$ (decayed by 0.1 every 30 epochs), batch size of 32, and total 100 epochs is used. An 80/10/10 data split with five-fold cross-validation is conducted, and consensus predictions are obtained via ensemble voting [(Feng et al., 2024), Sec. 3.1].

3. Grad-CAM Formulation for Visual Attribution

Grad-CAM (Gradient-weighted Class Activation Mapping) provides spatial heatmaps indicating which regions of an input most influence the model’s prediction for a specific class. For each forward pass, feature maps $A^k \in \mathbb{R}^{u\times v}$ at the last convolutional layer are weighted by gradients derived from the class score $I_{\text{masked}}$0:

• Gradients: $I_{\text{masked}}$1
• Channel weights: $I_{\text{masked}}$2 with $I_{\text{masked}}$3
• Combined map: $I_{\text{masked}}$4

After bilinear interpolation to input resolution, the resulting activation map can be superimposed on any ablated input, revealing task-specific cues under various masking regimes [(Feng et al., 2024), Sec. 3.2].

4. Quantitative Impact of Feature Ablation

Title classification results on the test set using the ensemble model demonstrate high discriminative capacity based solely on panel layout:

Input Type	Title Accuracy (%)
Unprocessed	87.5
Masked	79.7
Frame-only	84.3
Random Chance	~0.96

Comparable convergence is observed in learning curves for the original and frame-only inputs. For the subset of 4-panel works, accuracy reaches 93.1%. Publisher and genre classification with panel-only images yield macro-F1 ≈ 0.126 and overall accuracy of 20.8%, respectively, with the highest genre F1 for the 4-panel subset (0.78) and near-zero values for most others. No formal statistical significance testing is reported, but performance gaps are substantive and consistent [(Feng et al., 2024), Tables 1–4; Fig. 4].

5. Qualitative Attribution: Grad-CAM Findings

Grad-CAM visualizations differentiate attention strategies across ablation types:

• Unprocessed: Heatmaps primarily localize around character regions.
• Masked: Attention shifts toward panel borders, gutter spacing, and inter-panel alignment.
• Frame-only: Distinctive spatial arrangements are highlighted (e.g., "LoveHina’s" unique small bottom-right panel), suggesting that CNNs leverage topological relationships—size, spacing, and alignment—as robust work-specific “signatures.”
• 4-panel manga: Grad-CAM emphasizes inter-panel gutters, panel aspect ratios, and spatial distributions, capturing subtle author-specific layout conventions.
• Noisy frame variants: Random rectangular noise degrades classification by disrupting salient panel forms (e.g., bottom-right protrusion), while warped quadrilaterals preserve enough invariant structure for partial accuracy retention [(Feng et al., 2024), Figs. 5–7, 12].

6. Theoretical and Practical Implications

The results demonstrate that panel layouts, isolated from characters and text, are sufficiently distinctive for high-accuracy title recognition, challenging assumptions about the primacy of depicted content in stylistic attribution. Publishers and genres exert minimal influence on panel structure; instead, per-author or per-work variation dominates, as evidenced by the network’s orientation to spatial “signatures” even under spatial perturbations. Ablation CAM thus provides a robust methodology for dissecting the interplay between layout and content in visual media, with implications for interpretability, dataset design, and transfer to other structured document genres. Suggested future directions include extension to broader corpora, analysis of intra-author variability, and deployment of ablation plus CAM pipelines for interpretability in diverse document analysis contexts [(Feng et al., 2024), Sec. 4].