A multi-expert diffusion model for surface defect detection of valve cores in special control valve equipment systems

The valve core is a critical mechanical component of control valves, and accurate surface defect detection is essential for ensuring the reliable operation of special control valve systems. However, current methods often suffer from unsatisfactory detection accuracy due to the uncertain defect size and the strong interference of background grayscale. Additionally, many existing techniques depend on heuristic object priors (e.g., anchor boxes and reference points), reducing adaptability to complex environments. To tackle the above issues, a Multi-Expert Diffusion Model is proposed, integrating a Multi-Expert Feature Extraction (MFE) module, a Low-Pass Guided Feature Aggregation (LGFA) module, and a Heterogeneous Diffusion Detection (HDD) mechanism. Specifically, at the feature level, MFE explores the large kernel convolution and multi-expert selection mechanism, which ensures the model can dynamically adjust larger spatial receptive fields and better extract representative features of multi-scale defects from complex backgrounds. LGFA predicts spatial changes through a low-pass filter, reducing intra-class differences in the multi-scale feature fusion process and providing a complete feature basis for decision-making. At the decision level, HDD transforms defect detection into a denoising process that generates category distributions and bounding boxes, which weakens the model's reliance on heuristic object priors and ensures its scalability in detecting miscellaneous defects. Overall, the proposed method achieves high-accuracy surface defect detection from a novel generative perspective. Experiments on a real control valve platform show a 6.1% detection accuracy gain over special methods, providing a solid guarantee for the safe operation of electromechanical equipment.