A Deep Extractor for Visual Rail Surface Inspection

Rail surface inspection using a visual system is an important task in the maintenance of railway networks. Recent approaches have attempted to harness low-level features identify defects. These methods' main issues are limitations and prior information about rail surfaces that show wide variety appearances on dynamic backgrounds. To overcome these problems, we propose deep extractor (DE) combines strengths fully convolutional networks conditional random fields (CRFs), so network can learn abstract needed for specified task. Specifically, motivated by networks, bilateral containing two branches proposed here: one encode-decode branch focuses semantic meaning; other decode-encode encodes tiny The aggregated obtain high-level feature map. Furthermore, mean-field inference dense CRF with Gaussian pairwise potentials formulated as recurrent neural (CRF-RNN) introduced contribute smoothing constraints obtaining fine-grained result. By doing so, it avoid offline post-processing further lead whole architecture achieve end-to-end learning. Compared classical approaches, our approach outperforms state-of-art analyzing publicly available datasets.