Self-Supervised Learning of Free-Hand Sketches with Bézier Curve Features

Automatic sketch recognition is a challenging problem due to the abstract nature of the sketches, differences in artists’ abstraction and keystrokes. CNN-based methods alone are not enough to recognize abstract sketches as they do not have such a rich content as photographic images. In this study, we propose a method based on the Bezier curve transformation of strokes and self-supervised learning to eliminate the noise caused by either the data acquisition or differences in artist abstraction in the recognition task. The proposed method has been evaluated on a large-scale QuickDraw performance dataset and compared with state-of-the-art studies. The proposed method improves top-1 performance on the specific classes of QuickDraw dataset.