【摘要】 针对爆破片表面检测中多尺度缺陷敏感性不足及复杂纹理干扰较强的问题，提出融合空间动态加权与轴向长程注意力的多尺度协同增强方法。首先,提出空间动态加权机制，通过并行聚合均值与最大值特征生成空间权重图，实现对缺陷区域的像素级动态聚焦；其次，提出轴向长程注意力机制，采用分离式水平/垂直方向深度可分离卷积构建长程依赖建模路径，有效提升对微小缺陷的表征能力；最后，提出双路径特征表示机制，通过局部与全局信息联合进行特征表示与重建，增强模型泛化性。在MVTec-AD和自建爆破片数据集上的试验结果表明，该方法在缺陷检测与定位准确率上显著优于DRAEM、RD4AD和PatchCore等主流方法。更多还原

【Abstract】 To address the issues of insufficient sensitivity to multi-scale defects and interference from complex textures in the surface inspection of bursting discs, a multi-scale synergistic enhancement method was proposed, integrating spatial dynamic weighting and axial long-range attention. Firstly, a spatial dynamic weighting mechanism was introduced, which generated a spatial weight map by parallel aggregation of average and maximum features, enabling pixel-wise dynamic focusing on defect regions. Secondly, an axial long-range attention mechanism was proposed, utilizing separated horizontal/vertical depthwise separable convolutions to construct a long-range dependency modeling path, effectively enhancing the characterization capability for minute defects. Furthermore, a dual-path feature representation mechanism was presented, which jointly leveraged local and global information for feature representation and reconstruction, strengthening the model’s generalization capability.Experiments conducted on the MVTec-AD dataset and a self built bursting disc dataset demonstrated that this method significantly outperformed mainstream approaches such as DRAEM, RD4 AD, and PatchCore in terms of defect detection and localization accuracy.更多还原