Reconstructive Training for Real-World Robustness in Image Classification

David Patrick, Michael Geyer, Richard Tran, Amanda Fernandez

Abstract: In order to generalize to real-world data, computer vision models need to be robust to corruptions which may not generally be available in the traditional benchmark datasets. Real world data is diverse and can vary over time - sensors may become damaged, environments may change, or users may provide malicious inputs. While substantial research has focused separately on processing specific image distortions or on defending against types of adversarial attack, some real-world applications will require vision models to generalize to corruptions, while additionally maintaining image quality. We propose a simple training strategy to leverage image reconstruction, with similarities to a GAN training process, to reduce image data corruptions while maintaining the visual integrity of the image. Our approach is demonstrated on several corruptions for the task of image classification, and compared with established approaches, with qualitative and quantitative improvements. Code available at: https://github.com/UTSA-VAIL/ReconstructiveTraining