Using Collaged Data Augmentation to Train Deep Neural Net with Few Data

Hsun-An Chiang, Cheng-Shao Chiang, Chi-Sheng Shih

Keywords: endoscopy image, data augmentation

Abstract:

Image-guided surgery has shown its effectiveness on both open surgeries and minimal invasive surgeries in last several decades. In the meantime, augmented reality is designed to provide not only better user experience, compared to current image-guided systems, but also more dynamic information in real-time during the operations. However, several fundamental and critical challenges remain open. In this paper, we aim at semantic segmentation on endoscopy images to identify deformable objects and surgical instruments at pixel level. Specifically, this paper proposes a combined data augmentation approach to increase the amount of labeled endoscopy images for more than 16 times so as to enhance the learning results for objects of interests. Unlike other approaches, the combined data augmentation approach synthesizes the new labeled
images based on the type of the objects in the images. Experiment results show that data augmentation can increase mIoU
for up 14.6% for PSPNet and up to 39.83% for DeepLab networks.

Code Of Conduct: I have read and accept the code of conduct.

Remove If Rejected: Remove submission from public view if paper is rejected.