Overcoming Noise Interference and Sample Scarcity in Digital Radiography: A Synergistic Approach for Intelligent Pipeline Girth Weld Defect Recognition

Targeting the dual challenges of intense noise interference and limited labeled data in pipeline girth weld inspection, this study explores the intelligent recognition mechanism of Digital Radiography images. To address these pain points, a synergistic recognition framework is proposed, integrating adaptive denoising, collaborative enhancement, and geometry-adaptive detection networks. An intelligent noise reduction method is designed to solve the problem of high image noise. To address the poor quality of images, an image enhancement algorithm is proposed. A collaborative Fourier Transform is used to enhance the image through different channels. Finally, an improved YOLOv8-MSDCN model is constructed by combining the model of MSDCN with YOLOv8. This model realizes the coupling between geometric deformation adaptation and channel feature enhancement. The relationship between the position and number of MSDCN modules is investigated. It is found that when three MSDCN modules are embedded in the backbone, the model has the best recognition, with a precision of 0.87, mAP@0.5 of 0.86, and recall of 0.81. This proves that the integration of adaptive preprocessing and the fused detection mechanism significantly enhances the reliability of DR defect recognition.