Effect of Self-Referential Linear Processing on Deep-Learning-based Image Classification

抄録

Linearly mixing or combining multiple images is a frequently used image processing methods in computer vision. Mixup, which is a kind of linear operations, shows its effectiveness on improving the performance of deep-learning-based models and increasing the robustness of trained models against adversarial attacks. However, the effect and the underlying mechanism of linear operations are little understood. In this study, we investigate the effect of linear operations on the task of image classification. We apply several self-referential linear-mixing operations to process images, and use these images to evaluate the performance of deep-learning-based image classifiers under different mixing parameters. The contribution of this study is on establishing a foundation to better understand the underlying mechanism of linear operations.