Deep neural networks based approach for welded joint detection of oil pipelines in radiographic images with Double Wall Double Image exposure

Abstract

This paper describes a method to support the field of Nondestructive Testing, especially, in radiographic inspection activities. It aims at detecting welded joints of oil pipelines in radiographs with Double Wall Double Image exposure. The proposed approach extracts information (windows of pixels) from the pipeline region in the radiographic image and then applies Deep Neural Network (DNN) models to identify which windows correspond to welded joints. We use pre-trained DNNs to map the knowledge from ImageNet Large Scale Visual Recognition Challenge to the welded joint context. The experiments consider 13 DNN models and 3 DNN input settings: stretched, proportional V and proportional H. Since, occasionally, radiographic images may be corrupted by some types of noise (e.g. white, impulsive), we also include experiments considering its influence on the DNNs behavior and its related results. The best combination provided an F-score average of 96.00% in the welded joint detection. The main contributions of this work are the proposed window extraction technique and a robust analysis of the noise influence on welded joint detection using different DNN models, input settings, exposure techniques and radiographic acquisition sources.