Weak Appearance Aware Pipeline Leak Detection Based on CNN–Transformer Hybrid Architecture

Autonomous defect detection is critical for intelligent industrial production lines. Pipeline leakage as a typical defect presents an irregular or inconspicuous appearance, resulting in the failure of general visual inspection approaches in actual applications. To this end, we propose a contactless pipeline leak detection system with a novel semantic segmentation network, namely, WDA-Net, that can work in various illumination conditions. A novel convolutional neural network (CNN)–Transformer hybrid architecture in the encoder is designed to capture defect features, especially for features with weak appearance, while an efficient fusion strategy that integrates adjacent feature maps and gradually recovers the spatial details by skip connections is proposed in the decoder. Except for the categories and intrinsic characteristics of defects, complex illumination conditions have been considered in the constructed dataset. Comprehensive experiments demonstrate that the proposed method outperforms other methods, achieving 70.0% mean intersection over union (mIoU) on the pipeline leakage defect (PLD) dataset and 84.6% mIoU on the NEU-Seg dataset. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/zbl929/WDA-Net</uri>.