Texture Bias

Texture Bias is the tendency of convolutional neural networks to classify images primarily based on local texture patterns rather than global shape — CNNs rely on texture (surface patterns, colors, local statistics) more than shape (outlines, contours, global structure), while humans rely primarily on shape.

Texture Bias Evidence

• Stylized ImageNet: When texture and shape conflict (elephant shape with cat texture), CNNs classify by texture ("cat"), humans classify by shape ("elephant").
• Conflict Experiments: Geirhos et al. (2019) systematically showed CNNs use texture cues over shape cues.
• Robustness: Texture-biased models are less robust to distribution shifts — textures change more than shapes across domains.
• Training: Training on stylized images (removing texture) shifts CNNs toward shape bias and improves robustness.

Why It Matters

• Robustness: Shape-biased models are more robust to noise, domain shifts, and perturbations than texture-biased models.
• Alignment: Human perception is shape-based — aligning model features with human perception improves interpretability.
• Semiconductor: Defect classification should be based on shape/morphology, not texture artifacts from imaging conditions.

Texture Bias is judging by the surface — CNNs' preference for local texture over global shape, causing brittle, non-robust classification.