Deep Segmentation: using Deep Convolutional Networks for Coral Reef pixel-wise Parsing

In this paper, we describe a deep-convolutional network based method to segment coral reef images into different types of substrates. The method described in the paper includes data preparation, model summary, specific techniques to deal with class imbalance, and downstream post-processing computer vision tasks, such as morphological operations and polygon generation from pixel segmentation. We present the results of our method in the ImageCLEFcoral pixel-wise parsing task, evaluated across the different classes of substrate.