TreeSegNet: Automatically Constructed Tree CNNs for Subdecimeter Aerial Image Segmentation
For the task of subdecimeter aerial imagery segmentation, the fine-grained semantic segmentation results are usually difficult to obtain because of complex remote sensing contents and optical conditions. In addition, remote sensing imagery has inherent limitations of imbalanced class distribution. Recently, convolutional neural networks (CNNs) have shown outstanding performance on this task. In this paper, we propose the TreeSegNet to solve the class imbalance problem and further improve the accuracy in the metrics' point of view. Based on the infrastructure of DeepUNet, a Tree-CNN model in which each node represents a ResNeXt unit is constructed automatically according to confusion matrix and minimum graph cut algorithm. By transporting feature maps by concatenating connections, the Tree-CNN block fuses the multiscale features and learning the best weights for the model. In the experiments on ISPRS 2D semantic labeling Potsdam dataset, the results gotten by TreeSegNet are better than the opened state-of-the-art methods. The F1 measure scores of classes are improved especially for those classes that are easily confused. Completely and detailed comparison and analysis are performed to show that the improvement is brought by the construction and the embedding of the Tree-CNN module.
READ FULL TEXT