TransVJ2-MCA: power line defect recognition integrating video world model priors and cross-attention

To address the dual challenges of scarce defect samples in power line component images and fine-grained defects being easily overwhelmed in complex backgrounds, this paper proposes a video world model prior-driven component defect recognition framework, TransVJ2-MCA. First, a cross-modal temporal reconstruction (CMTR) mechanism is designed to adapt to the input constraints of the video self-supervised pre-trained backbone network (V-JEPA v2). By replicating single-frame power line component images along the temporal dimension and converting them into stationary pseudo-video sequences, static images are aligned with the V-JEPA v2 encoder in both input form and feature distribution. A parameter-freezing strategy is then adopted to reuse the large-scale spatio-temporal structural priors embedded in the model, mitigating the limitation imposed by scarce defect samples on model training. Finally, a multi-scale cross-attention (MCA) module is constructed, which uses deep semantic features as guidance to selectively aggregate key discriminative information from shallow features, enabling effective selection and adaptive fusion of cross-level features, suppressing complex background interference, and enhancing fine-grained defect recognition capability. Experimental validation is conducted on five independent power line component classification tasks of the InsPLAD-Fault dataset. The results show that TransVJ2-MCA achieves an average balanced accuracy of 98.2% and a macro-average F1 score of 97.1%, demonstrating the effectiveness of the proposed method under complex backgrounds and scarce defect samples, as well as its cross-modal transfer capability. This study provides a referential adaptation paradigm for transferring video pre-trained large models to static industrial visual inspection.