Quantifying and Inducing Shape Bias in CNNs via Max-Pool Dilation
Takito Sawada, Akinori Iwata, Masahiro Okuda
TL;DR
We address the CNN texture bias by introducing a data-driven $L0$-SSIM metric that measures the shape–texture balance of a dataset via the $SSIM$(Y,$L0$) between the luminance channel $Y$ and its $L0$-smoothed version; a higher score indicates shape dominance. We then propose an efficient adaptation that freezes convolutional weights and shifts bias by changing max-pooling dilation from $1$ to $2$, training only the final classifier. On six small datasets, the metric-guided approach selects the appropriate bias: shape-dominant datasets benefit from a shape-biased $S_{conv}$-Model, whereas texture-dominant datasets favor the baseline, with max-pool dilation offering selective gains but potential losses, validating a metric-guided approach; additionally, the max-pool dilation method yields modest gains for shape-dominant data but can hurt texture-rich data, highlighting the need for dataset-aware deployment. Overall, the framework enables efficient, data-driven bias alignment in low-data regimes.
Abstract
Convolutional Neural Networks (CNNs) are known to exhibit a strong texture bias, favoring local patterns over global shape information--a tendency inherent to their convolutional architecture. While this bias is beneficial for texture-rich natural images, it often degrades performance on shape-dominant data such as illustrations and sketches. Although prior work has proposed shape-biased models to mitigate this issue, these approaches lack a quantitative metric for identifying which datasets would actually benefit from such modifications. To address this gap, we propose a data-driven metric that quantifies the shape-texture balance of a dataset by computing the Structural Similarity Index (SSIM) between each image's luminance channel and its L0-smoothed counterpart. Building on this metric, we further introduce a computationally efficient adaptation method that promotes shape bias by modifying the dilation of max-pooling operations while keeping convolutional weights frozen. Experimental results show that this approach consistently improves classification accuracy on shape-dominant datasets, particularly in low-data regimes where full fine-tuning is impractical, requiring training only the final classification layer.