Breaking the "Stitch-then-Detect" Paradigm: A Real-Time AI Methodology for Synergistic Flaw Identification in Complex X-ray Images

The safe operation of the transmission line is significantly influenced by the structural integrity of the tension clamp, a critical component of the power system. This investigation introduces a real-time approach to defect detection and merging in X-ray images that employs Swin Transformer and EfficientNet. The objective is to improve the intelligence and efficacy of tension clamp detection. The system is composed of two primary modules: The encoder-decoder architecture of E-RSNet is constructed using EfficientNet, which incorporates a multi-head self-attention mechanism and recursive feature fusion to achieve high-precision image stitching. The Swin Transformer is the foundation of E-DDT, which incorporates a multi-scale attention mechanism and an adaptive feature pyramid network to improve the recognition capabilities of a variety of minor defects. Attention-guided feature fusion and multi-task learning frameworks enable the system to optimize image stitching and defect detection tasks in a coordinated manner. Lightweight methodologies, such as dynamic pruning and mixed precision quantization, are implemented to satisfy the real-time requirements of on-site detection. A prototype system was constructed and experimentally validated to confirm the end-to-end feasibility and real-time performance of the proposed framework. The proposed method outperforms existing techniques in image stitching quality, defect detection accuracy, and computational efficiency, as demonstrated by empirical data validation and simulation experiments. This approach is a viable option for the intelligent and portable detection of electrical equipment.