TLDR: Researchers propose two regularization methods to make Convolutional Neural Networks (CNNs) more robust to image corruptions by encouraging them to focus on global object shapes rather than local textures. The first method uses frequency filtering to make features consistent between original and low-frequency inputs, while the second employs supervised contrastive learning to group similar class representations. Both strategies improve CNN performance on corrupted images without sacrificing accuracy on clean data, moving CNNs closer to human-like perception.
Convolutional Neural Networks, or CNNs, have achieved remarkable success in tasks like image classification, powering everything from facial recognition to medical diagnostics. However, despite their impressive accuracy on clean datasets, these powerful AI models often falter when confronted with common image corruptions such as noise, blur, or fog. This fragility is a significant hurdle for deploying CNNs in real-world applications where input data is rarely perfect.
Understanding CNN Vulnerabilities
A key reason for this vulnerability lies in how CNNs ‘see’ the world. Unlike humans, who primarily rely on global object shapes for recognition, standard CNNs tend to focus on local texture patterns. This texture bias makes them highly susceptible to corruptions that primarily distort high-frequency components – the textures – while leaving the underlying shape information relatively intact. Imagine trying to identify a cat solely by the pattern of its fur; if the fur is blurred or noisy, identification becomes much harder, even if the cat’s overall form is still discernible.
Two Novel Approaches to Enhance Robustness
To address this fundamental limitation, recent research proposes two complementary regularization strategies designed to encourage CNNs to develop a stronger ‘shape bias’ and, consequently, become more robust to corruptions. These methods modify the network’s training process by introducing auxiliary loss functions, guiding the optimization towards more resilient representations.
1. Frequency-Based Regularization
The first approach operates in the frequency domain, leveraging the insight that shape information typically resides in low-frequency components of an image, while textures are found in high-frequency components. This method introduces a custom ‘FrequencyFilterCNN’ layer that encourages the network to produce similar feature activations for both original images and their low-frequency filtered counterparts. By forcing consistency between these two versions, the network is biased towards learning features that remain stable even when high-frequency textures are suppressed, effectively promoting a focus on shape-relevant information. This is achieved by transforming the input to the frequency domain, applying a Gaussian low-pass filter (which blurs the image by removing sharp details), and then transforming it back to the spatial domain. The network is then trained to minimize the difference in activations between the original and the blurred image.
2. Supervised Contrastive Learning
The second strategy employs supervised contrastive learning, a technique that explicitly structures the network’s internal feature space. Instead of just focusing on classifying images correctly, this method pulls together the representations of samples belonging to the same class while pushing apart representations of samples from different classes. The hypothesis here is that shape is generally more consistent within a class (e.g., all cats share similar facial structures) than texture, which can vary widely. By encouraging the network to learn class-invariant features, this regularization naturally promotes a shape bias, as these stable features are more likely to be shape-based rather than texture-based.
Also Read:
- Enhancing Medical Image Alignment Through Decoupled Training Supervision
- Enhancing Audio Event Recognition Through Consistency Regularization
Experimental Validation and Promising Results
These innovative regularization methods were rigorously evaluated on the CIFAR-10-C benchmark, a challenging dataset that applies 19 different types of corruptions across five severity levels to the standard CIFAR-10 dataset. Using a ResNet-18 architecture as the backbone, both proposed methods demonstrated improved robustness against common corruptions without compromising accuracy on clean, uncorrupted images.
Notably, the supervised contrastive learning approach achieved the best overall performance, showing a 0.8% improvement in mean corruption robustness accuracy compared to the baseline model. The frequency-based regularization also contributed positively, improving robustness accuracy by 0.3%, and proved particularly effective against blur-related corruptions. These results suggest that by strategically modifying the loss function during training, CNNs can be effectively guided to develop more shape-aware and resilient representations, bridging the gap between artificial and human visual perception.
For a deeper dive into the technical specifics, you can read the full research paper here.
The findings underscore the potential of loss-level regularization as a powerful tool to enhance the practical deployment of CNNs in diverse real-world scenarios where data quality cannot always be guaranteed. Future work aims to further explore these methods, including evaluating them on stylized datasets to directly quantify shape and texture bias, and investigating the synergistic effects of combining both regularization techniques.
