Di LIU, Xiao-Chun HOU, Yan-Bo LIU, Lei Liu, Yan-Cheng Wang, University of Electronic Science and Technology of China, China

Object detection typically requires a large amount of data to ensure detection accuracy. However, it is often impossible to ensure sufficient data in practice. This paper presents a new data augmentation method based on pixel-level image blend and domain adaptation. This method consists of two steps: 1.Image blend using a labeled dataset as object instances and an unlabeled dataset as background images.2. Domain adaptation based on Cycle Generative Adversarial Networks (Cycle GAN).A neural network will be trained to transform samples from step 1 to approximate the original dataset. Statistical consistency between new dataset generated by different data augmentation methods and original dataset will be measured by metrics such as generator loss and hellinger distance. Furthermore, a detection/segmentation network for diabetic retinopathy based on Mask R-CNN will be built and trained by the generated dataset. The effect of data augmentation method on the detection accuracy will be presented.

Data Augmentation, Object Detection, Image Blend, Domain Adaptation, Diabetic Retinopathy