Visual Data Restoration for Pipeline Inspection Robots by Compressed Sensing

Abstract

Pipelines are the key infrastructure for gas and liquid fossil energy, any leakage on the transportation pipelines lead to disasters for human safety and environment. The most popular method for the health monitoring of it is using pipeline inspection robots. The circumferential-deployed sensors capture two-dimensional data, i.e. visual images for defect analysis. However, due to the high vibration and harsh environment in pipelines, sensor failure becomes the common cases. Fixing this sensor failure is solving one visual data restoration problem, so this paper proposes one compressed sensing (CS)-based method. The core idea is that sampling a partial pixels in one image can full reconstruct the full image with CS theory, and the health sensor values are regarded as the CS sampling values. Block-by-block restoration scheme is used. The experimental results on real pipeline inspection data show our proposed CS method has better performance than the traditional interpolation method for most typical features, e.g. 20% of improvement on 2D coefficients with baseline for spiral weld, 53% of improvement on root-mean-square-error on defect region.