Active Learning-Based Complex Pipeline Weld Defect Detection with Lightweight Neural Network

Abstract

Weld defect detection plays an important role in pipeline safety maintenance. Due to the complexity of weld defect, it is difficult for existing intelligent detection methods based on deep learning to achieve high accuracy and efficiency. We propose a novel defect detection method that can better alleviate the current dilemma by constructing an active learning framework with well-designed value sample sampling strategy. First, primary defect detector is trained based on data driven by building a lightweight fully convolutional network. Then efficient value sample selection strategy is devised by computing the uncertainty of unlabeled images. Finally, Fine-tuning network parameters based on value samples is proposed to obtain large performance gains with minimal resources. The experimental results on the pipeline weld defect dataset in northern China show that the proposed method has higher detection accuracy than traditional deep learning methods.