Deep stochastic image segmentation for autonomous robotic inspection

Abstract

Corrosion detection inside a vessel’s ballast tank is a dangerous task that is normally performed by human specialists who, due to standards that are not easily quantifiable, can cause a high level of subjectivity in the inspection process. The Inspectrone project aims to automate this process using a drone flying through the different ballast tanks to perform the inspection, avoiding possible risks for the crew. Due to the subjectivity of the task a standard, deterministic, model for semantic segmentation can not be used for this scenario. On the contrary, a model that performs stochastic segmentation is required to enable the encapsulation of multiple differentiating human specialist opinions. The architecture presented in this paper combines a segmentation model with a latent distribution to encode the segmentation variants of the dataset to solve this task by being able to produce, for the same input image, different, but all plausible segmentations. The designed model is defined as Probabilistic GSCNN since it employs the state-of-the-art Gated Shape CNN to perform segmentation and a fixed prior distribution to produce a probabilistic output. The proposed method was able to outperform the baseline method, the Probabilistic U-Net, with a 9% increase in accuracy and without the need to train the CVAE module allowing a faster and less resource-demanding development.