Small Object Segmentation Using Dilated Convolutions With Increasing-Decreasing Dilation

This article presents a novel convolutional neural network (CNN) architecture for segmenting significantly small and crowded objects in remote sensing imagery. Although such small objects are characteristic in the remote sensing domain, the previous works mostly follow the state-of-the-art CNN models designed for ground-based images and have yet to fully explore the method for segmenting the small objects. To this end, we propose a network with no downsampling layers by utilizing dilated convolutions. We find that naive use of dilated convolutions with "increasing" dilation rates fails to capture local relationships among neighboring features, resulting in grid-like noise in the prediction. To alleviate this problem, we propose a novel scheme of "increasing-decreasing" dilation rates. Specifically, we propose a network module with decreasing dilation rates and attach it to the dilated backbone to reconnect the neighboring pixels of the backbone features. In the experiments, we evaluated the proposed model on six remote sensing datasets, where the model showed remarkably high performance, especially for small objects.