Shape Bias in CNNs: Metrics, Methods, and Robustness

• Shape bias in CNNs is defined as the model’s tendency to rely on global object contours instead of local textures for image classification, paralleling human vision.
• Empirical findings show that standard CNNs are largely texture-biased, while interventions like stylized data training and edge augmentation can significantly shift bias toward shapes to improve robustness.
• Algorithmic strategies, including loss-level regularization and data-centric augmentations, effectively promote shape bias, enhancing performance under adversarial attacks and domain shifts.

Shape bias in Convolutional Neural Networks (CNNs) refers to a model’s reliance on global object contours or outlines, as opposed to local textures, when making classification decisions. Human visual systems are strongly shape-biased, systematically favoring overall form in recognition tasks, whereas standard CNNs trained on natural images typically exhibit a marked texture bias. Researchers have characterized, quantified, and manipulated shape bias through psychophysics-inspired benchmarks, architectural interventions, loss-level regularization, and data-centric augmentations. The presence and strength of shape bias in CNNs are linked to robustness under distribution shift, adversarial perturbation, and domain generalization performance (Geirhos et al., 2018, Mohla et al., 2020, Li et al., 2023).

1. Definitions, Metrics, and Human Comparison

Shape bias is formally defined as the classifier's propensity to select class labels based on global shape rather than local texture in cases where these cues are in conflict. Foundational psychophysical protocols introduce cue-conflict stimuli: images that have the silhouette (shape) of one object and the texture of another. The shape bias index is typically defined as: $$\text{ShapeBias} = \frac{N_{\rm shape}}{N_{\rm shape}+N_{\rm texture}},$$ where $N_{\rm shape}$ denotes the number of shape-consistent responses on cue-conflict images and $N_{\rm texture}$ denotes texture-consistent responses (Geirhos et al., 2018, Hermann et al., 2019). Human participants exhibit high shape bias ($\approx$0.96), consistently prioritizing global form (Geirhos et al., 2018).

A related metric is the shape bias score under negative (brightness-inverted) images: $$S_{\text{bias}} = \frac{\mathrm{Acc}_{\rm neg}}{\mathrm{Acc}_{\rm orig}},$$ where $\mathrm{Acc}_{\rm orig}$ and $\mathrm{Acc}_{\rm neg}$ are the classification accuracies on original and inverted images, respectively (Hosseini et al., 2018).

Further approaches include direct per-layer representational similarity analysis (RSA) comparing layer activations’ alignment to shape or texture labels (Islam et al., 2021), neuron-level correlation tests (Iwase et al., 4 Mar 2025), or the Disrupted Structure Testbench (DiST), which quantifies sensitivity to global structure independently of local statistics (Wen et al., 2023).

2. Empirical Findings: Inductive Biases of Standard CNNs

Empirical work demonstrates that standard ImageNet-trained CNNs are highly texture-biased: on cue-conflict stimuli, vanilla ResNet-50 yields shape bias $\approx 0.22$, indicating 78% of correct answers follow the texture rather than the shape (Geirhos et al., 2018). This pattern is robust across architectures and even persists under most unsupervised objectives, with texture decisions dominating even in deeper models (Hermann et al., 2019). Human observers, in contrast, virtually always select by shape on these tasks.

Examination of latent representations via RSA shows that shape information is largely confined to the late convolutional layers, and that even when these layers are shape-sensitive, they do not necessarily encode per-pixel semantic detail—CNN global shape sensitivity does not entail explicit segmentation (Islam et al., 2021).

3. Determinants of Shape Bias: Data, Augmentation, and Initialization

Data-centric factors, especially the nature and diversity of the training set and applied augmentations, have a marked influence on shape bias. Standard data augmentation strategies such as random cropping, color distortion, Gaussian blur, and noise collectively can shift a ResNet-50 from 19.5% to over 60% shape bias, with the most pronounced effect coming from increased minimum crop size and the addition of naturalistic augmentations (Hermann et al., 2019). Using negative images—where shape is preserved while brightness is inverted—as an evaluation and augmentation tool serves as a practical proxy for shape invariance (Hosseini et al., 2018).

Initialization and regularization further modulate shape bias: He/Xavier initializations, batch normalization, and geometric plus color/brightness perturbations all significantly boost the shape bias ratio in different architectures (Hosseini et al., 2018).

4. Algorithmic Strategies to Promote Shape Bias

Several methodologies have been proposed to counteract intrinsic texture reliance:

• Stylized Data Training: Training on "Stylized-ImageNet," where original textures are replaced by artistic styles via adaptive instance normalization (AdaIN), shifts the same architecture (ResNet-50) from texture-bias (shape bias ≈ 0.22) to strong shape-bias ($\approx$ 0.81) and confers robustness on out-of-distribution distortions (Geirhos et al., 2018).
• Domain-Adversarial Learning: Adding a domain-adversarial head to distinguish between original and stylized inputs, and actively penalizing domain-distinguishing in feature space, further strengthens shape bias (Brochu, 2019).
• Edge-Augmentation and Shock Graphs: Feeding edge maps as input channels or using graph neural networks over shock graph–based shape representations (contour-only descriptors) enforces shape-centric classification and can outperform RGB baselines under heavy domain shifts (Borji, 2020, Narayanan et al., 2021).
• Loss-Level Regularization: Losses that penalize discrepancies between original and low-pass filtered versions of inputs (frequency-based), or that cluster feature representations by class with supervised contrastive learning (SupCon), encourage class-invariant, shape-aware features and improve robustness (Ranabhat et al., 14 Sep 2025).
• Sparsity and Top-K Activation: Non-differentiable spatial Top-K layers that select the K most active units per channel induce sparse, part-oriented responses and result in a marked increase in shape bias, even in off-the-shelf networks (Li et al., 2023).

5. Interventions and Their Effects: Robustness, Generalization, and Trade-offs

Shape-biased representations confer improved robustness to common corruptions, domain shift, and adversarial attacks, although the correlation between shape bias and corruption robustness is not always monotonic (Mummadi et al., 2021, Ranabhat et al., 14 Sep 2025). Stylized data augmentation and explicit edge information both enhance corruption robustness, but the effect size is highest when these interventions are combined with structural or frequency-level regularizers (Borji, 2020, Ranabhat et al., 14 Sep 2025).

Adversarially-trained networks (e.g., via PGD) exhibit much higher shape bias (70%+), with smoother filters and a reduction in the diversity and complexity of features activated by each neuron (Chen et al., 2020). Similarly, InfoDrop, which stochastically drops low self-information (texture-like) regions, produces networks that fail rapidly when patch shuffling destroys global shape, directly indicating greater shape reliance (Shi et al., 2020).

Methods such as CognitiveCNN explicitly enforce a match between the relevance of different modalities (shape, texture, edges, greyscale) during both reconstruction and classification, resulting in simultaneous boosts to clean accuracy and resilience to cue-conflict stimuli (Mohla et al., 2020).

6. Analytical Perspectives and Insights

Recent research reveals that shape bias is not an intrinsic property of CNN architectures but a malleable characteristic dependent on training objectives, data properties, and augmentations (Hosseini et al., 2018). Architectures with identical original-set accuracy can differ widely in their shape bias under negative or stylized test conditions.

Notably, shape bias and robustness to corruptions can diverge: models with high shape bias do not always exhibit the highest mean corruption accuracy, and resistance to style-transfer attacks does not guarantee sensitivity to global structure as measured by DiST (Wen et al., 2023). This decoupling indicates that enhancing shape bias and ensuring robustness are related but not synonymous tasks, and optimal performance is often achieved by combining multiple interventions (e.g., DiSTinguish training with stylized augmentation) (Wen et al., 2023). Monitoring the evolution of shape and texture bias during training reveals synchronization with the double-descent phenomenon, suggesting a dynamic allocation of feature bandwidth between global (shape) and local (texture) cues (Iwase et al., 4 Mar 2025).

7. Practical Guidelines and Future Directions

To build and diagnose shape-biased CNNs:

• Incorporate batch normalization and data augmentation (crop, flip, color/brightness shifts) as standard practices to suppress reliance on absolute pixel intensities (Hosseini et al., 2018).
• Use stylized images, negative images, edge-based input augmentations, or domain-adversarial training to promote shape-based inductive bias (Geirhos et al., 2018, Brochu, 2019, Borji, 2020).
• Deploy loss-level frequency regularization, supervised contrastive clustering, Top-K activation sparsity, and multi-modal attention regularization for more controlled bias tuning (Ranabhat et al., 14 Sep 2025, Li et al., 2023, Mohla et al., 2020).
• Monitor explicit shape bias scores (e.g., Acc_neg/Acc_orig, cue-conflict performance, patch-shuffle accuracy) in addition to standard accuracy during development and model selection (Hosseini et al., 2018, Hermann et al., 2019).
• Combine structure-sensitive training such as DiSTinguish with stylization for best results in both global structure sensitivity and robustness to local corruption (Wen et al., 2023).

Emerging lines of research connect shape bias modulation to fundamental learning curves (e.g., double descent), highlight the need for global and local bias disentanglement, indicate that sparse codes promote part-based and robust features, and suggest that loss-level regularization plus explicit shape-focused augmentations are promising levers for next-generation robust and human-aligned visual systems (Li et al., 2023, Iwase et al., 4 Mar 2025, Wen et al., 2023).