Building Footprint Extraction from Aerial Images using Multiresolution Analysis Based Transfer Learning

This paper proposes a multiresolution analysis based transfer learning method for building footprint extraction using the U-Net architecture. This involves pre-training the network on the first level approximation of the dataset obtained by using the wavelet transform before transferring the weights to a new model for fine-tuning on the original dataset. The intuition behind this stems from the fact that image data possesses features at different scales, thus training sequentially on these multiple resolutions should be beneficial. Experimental results show that the introduced transfer learning method provides a better segmentation performance and requires lesser training time when compared to other state-of-the-art weight initialization methods. We further analyse how the usage of different wavelets for decomposing the original dataset affects the task, with Haar giving the best performance from the set of wavelets considered in this study. The approach introduced in this work is shown to be an effective way of augmenting the U-Net with multiresolution information, without the need of making any changes in the standard architecture.