Enhancing Downstream Task Performance of Weld SEAM Images Using Improved Unsupervised Training Strategies

In the field of non-destructive testing (NDT), the complexity of weld seam images and the scarcity of defect samples lead to suboptimal performance of existing models in defect classification and feature extraction tasks. To address this issue, a novel method is proposed that aims to learn the capability of underlying feature extraction from weld seam images instances regions with distinctive texture characteristics, thereby enhancing model performance in downstream tasks. This method employs semi-supervised learning to obtain instance data from a large set of weld seam photographs, leveraging this data to learn the differences in texture features among the photographs. Concurrently, a hierarchical clustering approach for hard sample selection is introduced, optimizing the training process of contrastive learning by selectively choosing challenging hard samples, thereby enhancing the model's ability to learn fine-grained features. Experimental results demonstrate that the proposed method significantly improves model performance in downstream tasks such as defect classification and feature extraction compared to traditional approaches.