MobileNetV3 for Pipeline Welding Defect Detection: A Computer Vision Approach on Noisy RGB Images

Weld defect detection plays a central role in pipeline inspection because it directly influences structural reliability and operational safety.Conventional manual inspection methods suffer from several limitations, including high labor demands, subjective judgment, and inconsistent results, which restrict their effectiveness in large-scale industrial applications. To overcome these challenges, this study evaluates the use of the lightweight convolutional neural network MobileNetV3, considering both its Small and Large configurations, for automatic multi-class classification of pipeline weld defects.The proposed approach integrates image preprocessing and data augmentation with a lightweight deep learning architecture to extract discriminative visual features from weld images. Experimental validation is carried out using two datasets acquired under practical inspection conditions. The AI5083 dataset contains 4,927 images captured using HDR cameras, while the PWDID dataset includes 78 images collected with a mobile phone camera in uncontrolled environments.The experimental results demonstrate strong classification performance, achieving an accuracy of 99.99% on the AI5083 dataset and improved results on the PWDID dataset. Compared with previously reported methods, the proposed approach yields an accuracy improvement of approximately 3%, indicating its robustness across different acquisition settings. In addition to its high accuracy, MobileNetV3 offers fast inference, low memory consumption, and reduced power requirements. Owing to these characteristics and ias compatibility with deployment frameworks such as TensorFlow Lite, ONNX, and CoreML, MobileNetV3 is well suited for mobile and handheld inspection tools. Based on these findings, the use of MobileNetV3 and other modern lightweight architectures along with denoised function are recommended for practical, on device pipeline weld defect inspection systems.