In this paper, the authors present a technique to improve the segmentation accuracy of a deepUNet (a UNet with Res Connections). The idea is to better disentangle classes that the UNet easily get confused with. The proposed solution is called TreeSegNet which adopts an adaptive network to increase the classification rate at the pixelwise level.
The overall method is summarized in Fig.1. First, the method passes a deepUNet for segmentation. Then, based on the confusion matrix of the UNet, a treeNetCNN is build and appended right after it.
Building the CNN tree
The CNN tree is built with the following 4 steps:
- Calculate the lower triangular matrix;
- Build the undirected graph;
- Run the iterated the TreeCutting operation;
- Build the tree structure.
These operations are summarized in Fig.3. Note that each node of that tree is a ResNeXt block as shown in Fig.4
The intuition for the tree CNN is that the most easily confusing classes tend to choose the path that contains more neural layers and thus benefit from further feature extraction.
They tested their method on the well-known Potsdam remote-sensing dataset and outperformed state-of-the-art methods, although by a tiny margin.