A new scheme of ensemble deep models for corrosion detection in oil and gas structures

Abstract

Automatic visual inspection has gained a lot of attention last decades, thanks to the advancement in computer vision techniques such as deep learning. In oil and gas industry, keeping facilities in good serviceability is vital for the safety, environmental, production and cost . Corrosion is one of the threats that obviously reduces the mechanical proprieties of structures. Thus, without effective management of corrosion of the structures, rust of metal leads to catastrophic accidents, causing loss of lives and damages in environment and serious economic destruction. In this thesis, we present a new method for corrosion detection based on convolution neural networks (CNN). In particular, the major contributions of this thesis are: 1) we investigate the performance of multiple deep nets for the task of corrosion detection and we propose an ensemble scheme to strengthen the individual performance of the different deep nets, 2) we present a comprehensive literature review of corrosion detection works, 3) Introduce a new customized self-made dataset which is made up of corroded/non-corroded petroleum pipeline images, 4) we conduct an experimental evaluation for several methods from the state of the art as well as different handcrafted features, and 5) we investigate the effect of varying color space on the system performance. Experimental results proved the effectiveness of the proposed method. The proposed ensemble CNNs has outperformed several relevant state of the art methods as well as multiple handcrafted features.