Thanks to recent advances in CNNs, solid improvements have been made insemantic segmentation of high resolution remote sensing imagery. However, mostof the previous works have not fully taken into account the specificdifficulties that exist in remote sensing tasks. One of such difficulties isthat objects are small and crowded in remote sensing imagery. To tackle withthis challenging task we have proposed a novel architecture called localfeature extraction (LFE) module attached on top of dilated front-end module.The LFE module is based on our findings that aggressively increasing dilationfactors fails to aggregate local features due to sparsity of the kernel, anddetrimental to small objects. The proposed LFE module solves this problem byaggregating local features with decreasing dilation factor. We tested ournetwork on three remote sensing datasets and acquired remarkably good resultsfor all datasets especially for small objects.
展开▼
