Built-in Elastic Transformations for Improved Robustness
This work addresses robustness issues in computer vision for applications requiring reliable performance under natural image variations, representing an incremental improvement over existing methods.
The paper tackles the problem of neural visual classifiers lacking robustness to natural perturbations like elastic deformations, occlusions, and Gaussian noise by introducing elastically-augmented convolutions (EAConv), which improves general robustness on unseen perturbations and enhances performance on clean images without data augmentation, as demonstrated on CIFAR-10 and STL-10 datasets.
We focus on building robustness in the convolutions of neural visual classifiers, especially against natural perturbations like elastic deformations, occlusions and Gaussian noise. Existing CNNs show outstanding performance on clean images, but fail to tackle naturally occurring perturbations. In this paper, we start from elastic perturbations, which approximate (local) view-point changes of the object. We present elastically-augmented convolutions (EAConv) by parameterizing filters as a combination of fixed elastically-perturbed bases functions and trainable weights for the purpose of integrating unseen viewpoints in the CNN. We show on CIFAR-10 and STL-10 datasets that the general robustness of our method on unseen occlusion, zoom, rotation, image cut and Gaussian perturbations improves, while significantly improving the performance on clean images without any data augmentation.