Semantic segmentation is a difficult task, and even more difficult with limited data. In this paper, we employ two different approaches to semantically segment 30 images of dimension 1188×792 pixels. Our first method uses a convolutional neural network to train a classifier on subimages of 11×11 pixels. The second method involves training on a classifier on 20×20 pixel subimages, but using a different convolutional network of one convolutional layer with batch normalization and ReLU, a new deconvolutional layer and one loss layer. The first approach produces a somewhat coarser segmentation but fairly high accuracy, 68 percent. The second method, employing the new architecture, results in much crisper lines for a finer segmentation, but a lower accuracy.
展开▼
