Performing Particle Image Segmentation on an Extremely Small Dataset

Image segmentation is one of the typical computer vision tasks that has received great success with the recent advance of deep-learning methods. However, it is still a challenging problem, particularly when encountering limited data. In this paper, we present a new strategy for particle image segmentation, which relies on extensive data augmentation methods to reuse the available annotated samples for more effective performance. The procedure consists of the K-nearest neighbour (KNN) matting to fine-tune manually annotated boundaries and the use of data augmentations to expand available annotated data. After that, we employ the U-net architecture to train the model based on a small dataset consisting of twenty images. The results showed that the proposed strategy could effectively extract particle boundary features, thereby obtaining accurate segmentation results based on a very limited source of images.