Fine-Grained Tiny Defect Detection in Spiral Welds: A Joint Framework Combining Semantic Discrimination and Contrast Transformation

This is the authors' abstract. We don't add key points for this paper.

X-ray-based nondestructive testing (NDT) is crucial for quality monitoring and maintenance decision-making of spiral steel pipes. Deep neural networks have shown impressive results in weld defect detection. However, in real-world spiral welding systems, the differences in appearance and distribution between spiral weld images and other weld images lead to a decrease in the detection accuracy of these models. There are still challenges in accurately localizing tiny defects in pixel-redundant and low-contrast spiral weld images. Additionally, little attention has been given to the practical issue of the inherent blurring in spiral welding images due to X-ray imaging. To this end, we propose a joint framework combining semantic discrimination and contrast transformation (SDCT). Specifically, we design a semantic discriminative architecture that generates adaptively initialized features guided by semantics and alleviates defect scale differences. Subsequently, a contrast transformation mechanism (CTM) is introduced to dynamically optimize defect features. Finally, a progressive iterative decoding (PID) strategy is employed to generate fine-grained pixel-level outputs. Comprehensive experiments conducted on the newly constructed tiny spiral weld X-ray (T-SWX-ray) dataset and two public benchmarks demonstrate that SDCT achieves competitive performance using the state-of-the-art techniques, thereby providing a robust solution for fine-grained detection of tiny defects in spiral welds. The source code and T-SWX-ray dataset are publicly available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/SD-Weld/SDCT</uri>.