Classification With an Edge: Improving Semantic Image Segmentation with\n Boundary Detection

Abstract

We present an end-to-end trainable deep convolutional neural network (DCNN)\nfor semantic segmentation with built-in awareness of semantically meaningful\nboundaries. Semantic segmentation is a fundamental remote sensing task, and\nmost state-of-the-art methods rely on DCNNs as their workhorse. A major reason\nfor their success is that deep networks learn to accumulate contextual\ninformation over very large windows (receptive fields). However, this success\ncomes at a cost, since the associated loss of effecive spatial resolution\nwashes out high-frequency details and leads to blurry object boundaries. Here,\nwe propose to counter this effect by combining semantic segmentation with\nsemantically informed edge detection, thus making class-boundaries explicit in\nthe model, First, we construct a comparatively simple, memory-efficient model\nby adding boundary detection to the Segnet encoder-decoder architecture.\nSecond, we also include boundary detection in FCN-type models and set up a\nhigh-end classifier ensemble. We show that boundary detection significantly\nimproves semantic segmentation with CNNs. Our high-end ensemble achieves > 90%\noverall accuracy on the ISPRS Vaihingen benchmark.\n