Accurate segmentation of Ulva prolifera regions with superpixel and CNNs

Abstract

To get regions of Ulva prolifera, we propose a novel end-to-end way to segment the Ulva prolifera regions via aggregation of local classify prediction results. We creatively adopt SEEDS (Superpixels Extracted via Energy-Driven Sampling) to generate local multi-scale patches. We use powerful convolution neural networks to learn and classify the patches. At last, mapping the classify prediction results of patches to the whole image according to the patches classify prediction results, we can get more detailed segmentation of Ulva prolifera. As for the dataset, we collected images by UAV (unmanned aerial vehicle) in coastal waters off Qingdao, China. We show experimentally this method achieves great segmentation performance of Ulva prolifera, despite its indistinct features. In contrast, we train the model in fully convolutional networks for semantic segmentation based on our dataset, while our result achieves superior accuracy.