End to end real time deep learning framework for synergistic flaw detection in complex X ray images

The safe operation of transmission lines critically depends on the structural integrity of tension clamps, which are key load-bearing components in power systems. To address the challenges of incomplete X-ray field of view and reliable recognition of subtle internal defects, this paper proposes a real-time end-to-end framework for synergistic X-ray image stitching and defect detection based on EfficientNet and Swin Transformer. The proposed framework consists of two major modules. First, an EfficientNet-based encoder-decoder stitching network, termed E-RSNet, is developed by integrating multi-head self-attention and recursive feature fusion to achieve high-precision panoramic reconstruction from low-contrast X-ray images. Second, a Swin Transformer-based defect detection network, termed E-DDT, is constructed with an adaptive feature pyramid and multi-scale attention to enhance the localization and recognition of small and weak defects. In addition, a cross-task attention mechanism and uncertainty-weighted joint optimization strategy are introduced to enable coordinated learning between image stitching and defect detection. To satisfy on-site real-time requirements, lightweight optimization techniques, including structured pruning and mixed-precision quantization, are further incorporated into the framework. Experimental results demonstrate that the proposed method achieves superior performance in stitching quality, defect detection accuracy, and computational efficiency compared with representative baseline methods. A hardware prototype was also developed and validated, confirming the end-to-end feasibility and real-time performance of the proposed system in a field-deployable inspection setup. These results indicate that the proposed framework provides an effective solution for intelligent X-ray inspection of power line tension clamps.