Effective semantic pixel labelling with convolutional networks and Conditional Random Fields

TLDR

Deep convolutional neural networks have recently succeeded across many applications thanks to abundant training data and growing computing power. The study proposes a semantic pixel‑labeling method that combines CNN features, hand‑crafted features, and Conditional Random Fields. Per‑pixel class probabilities are generated from dense image patches using CNN and hand‑crafted features, then a Conditional Random Field refines the labeling by smoothing while preserving image edges. Applied to the ISPRS 2D semantic labeling challenge, the method achieves competitive classification accuracy.

Abstract

Large amounts of available training data and increasing computing power have led to the recent success of deep convolutional neural networks (CNN) on a large number of applications. In this paper, we propose an effective semantic pixel labelling using CNN features, hand-crafted features and Conditional Random Fields (CRFs). Both CNN and hand-crafted features are applied to dense image patches to produce per-pixel class probabilities. The CRF infers a labelling that smooths regions while respecting the edges present in the imagery. The method is applied to the ISPRS 2D semantic labelling challenge dataset with competitive classification accuracy.