Leveraging pre-trained Segmentation Networks for Anomaly Segmentation
Oliver Rippel, Dorit Merhof · 2021
Employing representations generated by large-scale training in a transfer-learning setting achieves state-of-the-art anomaly segmentation results when applied to the visual inspection task. Current approaches, however, focus exclusively on features of pre-trained classification networks, which are known to posess lower spatial resolution than segmentation or object detection networks. In our work, we investigate whether features extracted from pre-trained segmentation networks can be used to further improve anomaly segmentation performance in the transfer-learning setting. To this end, we apply state-of-the-art transfer-learning methods to encoder-decoder based segmentation networks. Results show that the encoders of pre-trained segmentation networks yield improved anomaly segmentation performance compared to their pre-trained classification counterparts. However, no consistent improvements can be observed yet regarding the decoders of the pre-trained segmentation networks. Together, this demonstrates that pre-trained segmentation networks can be used to further improve transfer-learned anomaly segmentation performance and that additional research is required to fully unleash their potential.