Patch Elastic Deformation: An Effective Data Augmentation Method

In image detection, various models and data enhancement algorithms have emerged in recent years. But these methods are either random copy-paste or rotation but seldom consider the local deformation caused by the partial extrusion of the object. Elastic deformation exists widely in the real world. Patch Elastic Deformation simulates the shape changes of things after various reality effects. As a complement to existing data augmentation methods, elastic deformation may appear anywhere in the sample, increasing the richness of the selection. Specifically, we randomly select a rectangle patch in the image and then apply elastic deformation to it, forcing models to extract input images’ essential features. For image classification and object detection tasks, we verify the effectiveness of our algorithm on CIFAR-10, CIFAR-100, and PASCAL VOC datasets. And we also compared it with the random erasing algorithm, proving that our algorithm can boost the performance of models.