Data Augmentation based on Invariant Shape Blending for Deep Learning Classification

In this article, we introduce a new technique for augmenting 2D shape datasets based on a planar blending. In particular, a recently introduced shape blending method is applied to numerous pairs of shapes extracted from a given class. Several in-between data belonging to the same category are therefore generated. While traditional data augmentation approaches mainly apply simple transformations to the original data, the proposed technique allows the generation of non-linear variations of the input shapes by covering significantly the shape space. To demonstrate its interest, our augmentation technique is applied to the task of 2D shape classification. Experiments are performed on two benchmarks, namely KIMIA'99 and MPEG-7 CE using two different Convolutional Neural Network (CNN) architectures. The results show the superiority of our method over traditional augmentation techniques.